Log in
The DDoS attack against Mastodon's flagship server comes less than a week after Bluesky was targeted with junk web traffic.
|
Sr. Content Developer at Microsoft, working remotely in PA, TechBash conference organizer, former Microsoft MVP, Husband, Dad and Geek.
|
Pennsylvania, USA
Remember when you had to really dig in concentrate and understand exactly how C# and other code worked at the most basic levels? Then you'll like Microsoft's early preview of .NET 11.
Pennsylvania, USA
The open-source Git project just released Git 2.54 with features and bug fixes from over 137 contributors, 66 of them new. We last caught up with you on the latest in Git back when 2.52 was released.
To celebrate this most recent release, here is GitHub’s look at some of the most interesting features and changes introduced since last time.
💡 Since the last Git release we wrote about was Git 2.52, this blog post covers the highlights from both the 2.53 and 2.54 releases.
Rewrite history with git history
The Git project has a long history of providing tools to rewrite your repository’s history. git rebase –i is the most well-known, and it’s remarkably flexible: you can reorder, squash, edit, and drop commits. But that flexibility comes with complexity: an interactive rebase operates on a range of commits, updates your working tree and index as it goes, and can leave you in a conflicted state that you need to resolve before proceeding.
For simpler cases, all of that machinery can feel like overkill. If all you want to do is fix a typo in a commit message three commits back, or split one commit into two, an interactive rebase works, but requires you to set up a to-do list, mark the right commit for editing, and then drive the rebase to completion.
Git 2.54 introduces a new experimental command that is designed for exactly these simpler cases: git history. The history command currently supports two operations: reword and split.
git history reword <commit> opens your editor with the specified commit’s message and rewrites it in place, updating any branches that descend from that commit. Unlike git rebase, it doesn’t touch your working tree or index, and it can even operate in a bare repository.
git history split <commit> lets you interactively split a commit into two by selecting which hunks should be carved out into a new parent commit. The interface will look familiar if you’ve ever used add in interactive mode via git add –p:
$ git history split HEAD
diff --git a/bar b/bar
new file mode 100644
index 0000000..50810a5
--- /dev/null
+++ b/bar
@@ -0,0 +1 @@
+bar
(1/1) Stage addition [y,n,q,a,d,p,?]? y After selecting hunks, Git creates a new commit with those changes as the parent of the original commit (which retains whatever hunks you didn’t select) and rewrites any descendent branches to point at the updated history.
There are a couple of intentional limitations worth noting. The history command does not support histories that contain merge commits, and it will refuse to perform any operation that would result in a merge conflict. By design, git history is meant for targeted, non-interactive rewrites, not the kind of open-ended history rewriting typically relegated to git rebase –i.
The history command is built on top of git replay‘s core machinery, which was itself extracted into a library as part of this work. That foundation means that git history benefits from replay‘s ability to operate without touching the working tree, making it a natural fit for scripting and automation in addition to interactive use.
This command is still marked as experimental, so its interface may evolve. Give it a try with git history reword and git history split, available in Git 2.54.
[source, source, source, source, source]
Config-based hooks
If you’ve ever wanted to share a Git hook across multiple repositories, you’ve probably had to reach for a third-party hook manager, or manually symlink scripts into each repository’s $GIT_DIR/hooks directory. That’s because, historically, Git hooks could only be defined as executable scripts living in one place: the hooks subdirectory of your .git directory (or whatever core.hooksPath points to).
That meant that if you wanted to run a linter before every commit across all of your repositories, you had to copy the script into each repository, which can be tedious and error-prone. Alternatively, you could set core.hooksPath to point to a shared directory, but that causes all of your repositories to share the exact same set of hooks, with no way to mix and match.
Git 2.54 introduces a new way to define hooks: in your configuration files. Instead of placing a script at .git/hooks/pre-commit, you can now write:
[hook "linter"]
event = pre-commit
command = ~/bin/linter --cpp20 The hook.<name>.command key specifies the command to run, and hook.<name>.event specifies which hook event should trigger it. Since this is just configuration, it can live in your per-user ~/.gitconfig, a system-wide /etc/gitconfig, or in a repository’s local config. That makes it straightforward to define a set of hooks centrally and have them apply everywhere.
Even better, you can now run multiple hooks for the same event. If you want both a linter and a secrets scanner to run before every commit, you can configure them independently:
[hook "linter"]
event = pre-commit
command = ~/bin/linter --cpp20
[hook "no-leaks"]
event = pre-commit
command = ~/bin/leak-detectorGit will run them in the order it encounters their configuration. The traditional hook script in $GIT_DIR/hooks still works, and runs last, so existing hooks are unaffected. You can see which hooks are configured (and where they come from) with git hook list:
$ git hook list pre-commit
global linter ~/bin/linter --cpp20
local no-leaks ~/bin/leak-detector Individual hooks can be disabled without removing their configuration by setting hook.<name>.enabled = false, which is particularly handy when a hook is defined in a system-level config but you need to opt a specific repository out.
Along the way, Git’s internal handling of hooks has been modernized. Many built-in hooks that were previously invoked through ad-hoc code paths (like pre-push, post-rewrite, and the various receive-pack hooks) have been migrated to use the new hook API, meaning they all benefit from the new configuration-based hook machinery.
Geometric repacking during maintenance by default
Returning readers of this series may recall our coverage of the new geometric strategy within git maintenance, which was introduced in Git 2.52. That strategy works by inspecting the contents of your repository to determine if some number of packfiles can be combined to form a geometric progression by object count. If they can, Git performs a geometric repack, condensing the contents of your repository without needing to perform a full garbage collection.
In 2.52, the geometric strategy was available as an opt-in choice via the maintenance.strategy configuration. In 2.54, it becomes the default strategy for manual maintenance. That means when you run git maintenance run without specifying a strategy, Git will now use the geometric approach instead of the traditional gc task.
In practice, this means that your repositories will be maintained more efficiently out of the box. The geometric strategy avoids the expensive all-into-one repacks that gc performs, instead combining packs incrementally when possible and falling back to a full gc only when it would consolidate the entire repository into a single pack. Along the way, it keeps your commit-graph, reflogs, and other auxiliary data structures up to date.
If you were already using maintenance.strategy = geometric in your configuration, nothing changes. If you hadn’t set a strategy (or were relying on the old gc default), you’ll start seeing the benefits of geometric repacking automatically. The gc strategy is still available if you prefer it and can be selected with maintenance.strategy = gc.
[source]
The tip of the iceberg…
Now that we’ve covered some of the larger changes in more detail, let’s take a closer look at a selection of some other new features and updates in this release.
The
git add –pcommand, Git’s tool for interactively staging individual hunks, received a handful of usability improvements in this release. When navigating between hunks with theJandKkeys, Git now shows whether you’ve previously accepted or skipped each hunk, so you don’t have to remember your earlier decisions.Separately, a new
--no-auto-advanceflag changes howgit add –phandles the transition between files. Normally, once you’ve made a decision on every hunk in a file, the session automatically moves on to the next one. With--no-auto-advance, the session stays put after you’ve decided on the last hunk, letting you use<and>to move between files at your own pace. This can be useful when you want to review your decisions holistically before committing to them.git replay, the experimental command for replaying commits onto a new base without touching the working tree, continues to mature. This release brings several improvements:replaynow performs atomic reference updates by default (instead of printingupdate-refcommands tostdout), has learned a new--revertmode that reverses the changes from a range of commits, can now drop commits that become empty during replay, and supports replaying all the way down to the root commit.Git’s HTTP transport now handles HTTP 429 “Too Many Requests” responses. Previously, a 429 from the server would be treated as a fatal error. Git can now retry the request, honoring the server’s
Retry-Afterheader when present, or fall back to a configurable delay via the newhttp.retryAftersetting. The newhttp.maxRetriesandhttp.maxRetryTimeconfiguration options provide control over how many times to retry and how long to wait, respectively.[source]
git log –L, which traces the history of a range of lines within a file, has historically used its own custom output path that bypassed much of Git’s standard diff machinery. As a result, it was incompatible with several useful options, including the-Sand-G“pickaxe” options for searching by content changes.This release reworks
git log –Lto route its output through the standard diff pipeline, making it compatible with patch formatting options and pickaxe searches for the first time.Say you want to trace the history of
strbuf_addstr()instrbuf.c, but only see commits wherelenwas added or removed within that function:$ git log -L :strbuf_addstr:strbuf.c -S len --oneline -1 a70f8f19ad2 strbuf: introduce strbuf_addstrings() to repeatedly add a string diff --git a/strbuf.c b/strbuf.c --- a/strbuf.c +++ b/strbuf.c @@ -316,0 +316,9 @@ +void strbuf_addstrings(struct strbuf *sb, const char *s, size_t n) +{ + size_t len = strlen(s); + + strbuf_grow(sb, st_mult(len, n)); + for (size_t i = 0; i < n; i++) + strbuf_add(sb, s, len); +}Prior to this release, options like
-S, (and-G,--word-diff, along with--color-moved) were silently ignored when used with-L. Now they work together naturally:-Lscopes the output to the function you care about, and-Sfilters down to just the commits that touched the symbol you’re searching for within it.[source]
Incremental multi-pack indexes, which we first covered in our discussion of Git 2.47 and followed up on in Git 2.50, received further work in this release. The MIDX machinery now supports compaction, which merges smaller MIDX layers together (along with their associated reachability bitmaps) to keep the number of layers in the chain manageable. This is an important step toward making incremental MIDXs practical for long-lived repositories that accumulate many layers over time.
[source]
git statuslearned a newstatus.compareBranchesconfiguration option. By default,git statusshows how your current branch compares to its configured upstream (e.g., “Your branch is ahead of ‘origin/main’ by 3 commits”). Withstatus.compareBranches, you can ask it to also compare against your push remote, or both:[status] compareBranches = @{upstream} @{push}This is useful if your push destination differs from your upstream, as is common in triangular workflows where you fetch from one remote and push to another (like a fork).
[source]
Say you have a series of commits, and want to add a trailer to each one of them. You could do this manually, or automate it with something like:
git rebase -x ‘git commit --amend --no-edit --trailer=”Reviewed-by: A U Thor <a href="mailto:<author@example.com>”’, but that’s kind of a mouthful.In Git 2.54,
git rebaselearned a new--traileroption, which appends a trailer to every rebased commit via theinterpret-trailersmachinery. Instead of the monstrosity above, we can now writegit rebase --trailer "Reviewed-by: A <a href="mailto:<a@example.com>"and achieve the same effect.[source]
When signing commits, a signature remains valid even when it was signed with a GPG key that has since expired. Previously, Git displayed these signatures with a scary red color, which could be misleading and lead you to interpret the signature itself as invalid. Git now correctly treats a valid signature made with a since-expired key as a good signature.
[source]
git blamelearned a new--diff-algorithmoption, allowing you to select which diff algorithm (e.g.,histogram,patience, orminimal) is used when computing blame. This can sometimes produce meaningfully different (and more useful) blame output, depending on the nature of the changes in your repository’s history.[source]
Under the hood, a significant amount of work went into restructuring Git’s object database (ODB) internals. The ODB source API has been refactored to use a pluggable backend design, with individual functions like
read_object(),write_object(), andfor_each_object()now dispatched through function pointers on a per-source basis. While none of this is user-visible today, it lays the groundwork for future features like alternative storage backends or more flexible object database configurations.git backfill, the experimental command for downloading missing blobs in a partial clone, learned to accept revision and pathspec arguments. Previously,backfillwould always download blobs reachable fromHEADacross the entire tree. You can now scope it to a particular range of history (e.g.,git backfill main~100..main) or a subset of paths (e.g.,git backfill -- '*.c'), including pathspecs with wildcards.This makes backfill much more practical for large partial clones where you only need historical blobs for a specific area of the repository.
[source]
Git’s alias configuration has historically been limited to ASCII alphanumeric characters and hyphens. That meant alias names like “hämta” (Swedish for “fetch”) or “状態” (Japanese for “status”) were off-limits. Git 2.54 lifts that restriction with a new subsection-based syntax:
[alias "hämta"] command = fetchThe traditional
[alias] co = checkoutsyntax continues to work for ASCII names. The new subsection form supports any characters (except newlines andNULbytes), is matched case-sensitively as raw bytes, and uses a command key for the alias definition. Shell completion has been updated to handle these aliases as well.The histogram diff algorithm received a fix for a subtle output quality issue. After any diff algorithm runs, Git performs a “compaction” phase that shifts and merges change groups to produce cleaner output. In some cases, this shifting could move a change group across the anchor lines that the histogram algorithm had chosen, producing a diff that was technically correct but visually redundant. Git now detects when this happens and re-diffs the affected region, resulting in tighter output that better matches what you would expect.
[source]
…the rest of the iceberg
That’s just a sample of changes from the latest release. For more, check out the release notes for 2.53 and 2.54, or any previous version in the Git repository.
The post Highlights from Git 2.54 appeared first on The GitHub Blog.
Pennsylvania, USA
Technical advancement is moving at such a rapid pace that it can be challenging to define the tomorrow we’re working toward. In The Shape of Things to Come, Microsoft Research leader Doug Burger and experts from across disciplines tease out the thorniest AI issues facing technologists, policymakers, business decisionmakers, and other stakeholders today. The goal: to amplify the shared understanding needed to build a future in which the AI transition is a net positive.
In this episode, Burger is joined by Amy Luers, head of sustainability science and innovation at Microsoft, and Ishai Menache, an optimization researcher at Microsoft Research, to explore how AI can both contribute to and help address climate change, emphasizing the need to separate hype from data and understand its real impact. While datacenters account for a small share of global emissions, their rapid growth raises local infrastructure concerns, even as AI offers powerful tools to optimize complex systems and accelerate climate solutions. The discussion frames AI as a critical but double-edged technology that must be steered carefully to support a sustainable future.
Subscribe to the Microsoft Research Podcast:
Transcript
[MUSIC]
DOUG BURGER: This is The Shape of Things to Come, a Microsoft Research Podcast. I’m your host, Doug Burger. In this series, we’re going to venture to the bleeding edge of AI capabilities, dig down into the fundamentals, really try to understand them, and think about how these capabilities are going to change the world—for better and worse.
In today’s podcast, I’m bringing in two experts to have a dialogue about the future of AI and sustainability. One, Amy Luers, is an expert on sustainability and the intersection of sustainability, technology, and science. And the other, Ishai Menache, is a world-renowned expert in optimization.
And so thinking about how technology can optimize systems, we’re going to talk about whether AI has the potential to help with climate change and sustainability and the degree to which there are challenges associated with AI. And we’re going to try to get to the root of the issue because that will determine the shape of things to come.
[MUSIC FADES]
I’m really excited about the two distinguished guests I have today. We have Amy Luers, who’s Microsoft’s senior global director for sustainability science and innovation. And we have Ishai Menache, who is a partner research manager at Microsoft Research.
And then the topic, of course, is AI and the climate and sustainability, which I think is on a lot of people’s minds. You know, we have a climate crisis happening. I’ve been a climate hawk since the 1990s. It’s something I, you know, worry a lot about. I care a lot about. Of course, Amy has devoted her career to it, so I can’t really talk, but it’s a really important issue.
And now, you know, we have this AI transition happening. We’re doing across the tech industry a large build-out of large, large, large computing systems, mega datacenters. And there’s a lot of concern in the world about how this might affect the climate, how … what this means.
Amy, we’ll talk, I think, about local communities, as well. And so I really wanted to dig in to the facts. What does this really mean? What do we think the impact’s actually going to be? Like, let’s separate the data from the hype and then also talk about some of the opportunities ahead because I do think there are things we’ll be able to do, and that’s why we have Ishai here.
So maybe I’ll first turn it over to Amy. Can you tell us a little bit about your job at Microsoft and what got you into this space, maybe a little bit of your story?
AMY LUERS: So as you said, I lead the sustainability science and innovation in the Microsoft Corp sustainability team, which really means I get to work with really smart people around the company, around the world, at MSR [Microsoft Research], on shaping and informing sustainability solutions for Microsoft but also for the world.
And part of that is leading our strategy on AI and sustainability. And how I got into it, I’ve been working on sustainability and climate my whole life. And I’ve worked, from the tech sector, I was at Google actually previously. Also was in the White House working at the intersection of the CTO’s office and environment and resources and energy.
I also led an international research institution, UN-based network rather, focused on sustainability. And in that context, after coming out of Google, where I was really … started to think about the power of compute and digital tools for transformation, and— which is why I was brought into the White House to work at that intersection—when I started leading the sustainability network, research network globally, Future Earth (opens in new tab), I really brought this need to think about innovation and digital technologies in that space.
And I will say the sustainability science network at that time, you know, it was 10 years ago, eight years ago maybe, was a little resistant to thinking about AI and technologies in this space.
And I started a global initiative called Sustainability in the Digital Age, where I really brought together the digital technology and AI community. It was in Montreal, which there’s a lot, a big AI community there, and the sustainability scientists globally. And really started to think about what are the potentials, what are the risks, and led a big international study to put together a research and innovation agenda in this space.
And that sort of really shifted my approach from just “big compute can help things,” which I was really focused on at Google, to this role of AI and machine learning in this space.
BURGER: And, Ishai, so you are a world-renowned expert in, now, ML and optimization. You know, you’ve published extensively. You’re, I think, famous in your research community. You’ve had, I think, broader—you’ve had a lot of impact on Microsoft’s business. You’ve also been published in the Harvard Business Review (opens in new tab).
So, you know, you’re a little bit polymathy and sometimes a little intimidating to me.
ISHAI MENACHE: Yeah. [LAUGHTER]
BURGER: You know, but I’d like to hear a little bit about your background, just, you know, a short version of your story for the listening audience.
MENACHE: Yeah. My background is actually in engineering. However, my graduate studies were, as you mentioned, like in ML, reinforcement learning, and later on distributed optimization, game theory, a little bit more on the theory side.
So my story is that when I was doing my postdoc at MIT [Massachusetts Institute of Technology], you know, the cloud was kind of on the rise, circa 2009 or so. And I got fascinated by the cloud. My initial interest was actually in the economics of the cloud and, you know, pricing. How you price the cloud. And I got to know about MSR because, you know, around that time there was a new kind of lab opening just by MIT, MSR New England. And I got fascinated by the cloud and, you know, not only the economic aspects of it, but more fundamentally, you know, how do you utilize resources more efficiently?
And that’s what got me to Microsoft Research in 2011. So I was consulting in MSR New England but then moved to Redmond in 2011 to join a lab called Extreme Computing group that was actually dealing with the cloud futures.
And if I can mention, Doug, you were also part of that, so …
BURGER: That’s right.
MENACHE: I’ve known you for quite some time.
BURGER: Yep.
MENACHE: And, you know, so sort of my, let’s say, my angle into that, so there were like a lot of systems people thinking about the, you know, infrastructure of cloud. And then at the other extreme, there were theoreticians. They were thinking about like, you know, the next kind of wave or like, you know, innovating in the area of algorithms.
But I think what was sort of missing is a little bit of bridging between, you know, algorithms and then cloud infrastructure. And that’s where I sort of found a very interesting niche for myself, and later on, for the group, which I founded in 2019.
BURGER: So you recently announced the system called OptiMind. And I, you know, I did a LinkedIn post (opens in new tab) about it because I was really excited about it.
And just tell us what the system does, like why … it got a lot of attention. So what does the system do? And then maybe we’ll dig a little bit into optimization for the audience. And then, but then we have to get back to AI.
MENACHE: For sure. So, you know, stepping back a little bit. So what is actually optimization or mathematical optimization? So optimization or mathematical optimization is a way of using mathematics to make the best decisions when there are many choices and some limitations. OK.
And, you know, just a little bit more, you know, concretely, so in every optimization problem you have, first of all, a description of the problem that you need to solve. Then you have a bunch of decisions that are actually, in mathematical terms, these are the variables. You have an objective. What is your goal? What are you trying to optimize?
It could be something that you’re maximizing, revenue, but it could be that you’re minimizing costs, so there are different versions or different kinds of goals or objectives. And then there are constraints, which is like you cannot do whatever you want. There are some sort of limitations such as capacity constraints in the cloud setting or other factors that you have to account for in order to come up with the best possible decisions.
BURGER: So what … so maybe a simple … just to be silly for a sec, so I have a, you know, complex drive to work, and one day I find that the way I usually take is blocked and my brakes are really worn and I can only stop twice. So your framework might be able to figure out, like, what path gets me there saving the most gas.
MENACHE: Right. So that’s one example. And maybe you don’t want to pay for tolls for some reason, so that limits, you know, the roads that you can take. You know, there’s speed limits and such things. These are all constraints that you have to account for.
BURGER: There might be some speed traps, but I’m willing to go by a speed trap if the route is much shorter.
MENACHE: Maybe.
BURGER: So stuff like that. So it gets pretty complicated, doesn’t it?
MENACHE: Right. It gets pretty complicated because especially when the, you know, maybe you’re a single driver, but in optimization settings, think of like some of the problems that we worked on actually with the Dynamics 365 was also in the context of field service, which is about managing technicians at scale. So think of like not just you, but thousands of technicians that have to fulfill or that have to take care of certain work orders. So it would be thousands or tens of thousands of work orders.
And then you need to assign the technicians to these work orders. And there’s a bunch of constraints. Maybe not every technician can do every work order. You have to account for the traveling of the technicians, right. So it’s like you’re not going to send a technician that is in Spokane to do something in, let’s say, in Seattle because, you know, all day will be wasted on traveling.
BURGER: And it’s not sustainable.
MENACHE: It’s not sustainable. [LAUGHTER] And also, you know, the gas, obviously. So all these kinds of considerations, you can map it formally into mathematical optimization. And then there are techniques of solving this problem to optimality.
So essentially there is some machinery and there are experts that can take these problems and come up with the algorithms, but not everyone can do it. So it requires some expertise. In fact, graduate-level expertise in operation research or in, you know, algorithms, computer science type of algorithms. And when gen AI was emerging, we saw an opportunity to democratize optimization with gen AI in the following sense that a person that is not an expert can define what they want to do.
You gave your example about, you know, getting to work. It could be like, you know, a simple example of packing, which is like I have a suitcase that I, you know, I have a limit of, like, 20 pounds. And I have a bunch of things that I have to … that I want to fit in like, you know, I have with certain importance. Some are more critical, like, you know, I don’t know, like my laptop, and all that. But then there are books that are quite heavy, and maybe I still want to read books.
BURGER: Or I’m running an airline, and I have to schedule the flights, and I want to minimize fuel.
MENACHE: Yeah, that, too. And essentially, so you want to be able to describe what you need to solve in plain English, specify the problem, say what the decisions are, like I mentioned, like what your goal is, and then what constraints need to be accounted for.
And you want to use AI that will help you take all these considerations and essentially formulate the algorithm itself. So write down the recipe, the mathematical recipe, that would produce an optimal solution.
BURGER: Got it.
MENACHE: So that’s what OptiMind is about. [It] is a small language model that was trained especially for this kind of scenarios of, you know, taking natural language and mapping it into an optimization algorithm.
BURGER: So this is really great, and I think we’re going to come back to this. I want to now go back to Amy. When we think about AI and these datacenters that the industry is building and they, you know, they use water, they use electricity, you know, there’s contention in some communities about them being placed there. If I, you know, I’d love to be really data-driven and just kind of very factual.
So if I look at the overall picture, like, what is the real impact we think of this transition on, you know, climate, sustainability. And it’s complicated, right, because there are many sources of emissions? Electricity gen is one, but you have renewable energy. But it takes materials to build these things. So can you kind of give us some framing to help us understand it?
LUERS: Yeah. So first of all, you know, I think when we think about AI and climate, a lot of people think about just the infrastructure side. And I think it’s really important to think about this holistically. I actually personally believe that AI will be one of the most influential factors determining our climate future, for better or worse. But I also believe that we actually need AI to solve the climate crisis. So with that as context, let’s talk about the infrastructure, remembering we have to really think about the full context. You know, let me put this into context.
So from a climate perspective, what matters is the emissions to the world, the emissions of greenhouse gases to the world, heat-trapping gases, …
BURGER: Right.
LUERS: … to the climate, not specifically energy, right, because energy can be in different forms.
BURGER: Right. It’s, what are you putting in the air?
LUERS: What are you putting in the atmosphere?
BURGER: That’s right.
LUERS: So, you know, if you think about it from a global perspective, the world uses about … energy itself accounts for about 75% of all of the emissions that go into the atmosphere (opens in new tab).
BURGER: Wow, that’s a lot.
LUERS: So that’s a lot. But a lot of people think it’s the whole thing. So there’s other things that are not energy. [LAUGHS]
BURGER: OK, three-quarters, three-quarters …
LUERS: But in the context of … so from a climate perspective, datacenters account for about .5, less than .5% of all emissions as of 2024 (opens in new tab).
BURGER: OK. But they’re growing?
LUERS: But they’re growing. And so if you’re growing and you think about … there are lots of projections. It’s hard to project really beyond a couple years, as you both know, because things are changing so quickly, both on the demand, on the efficiency, what we’re using. Are we going to be using small language models? Like, we don’t know what the future looks like.
The IEA [International Energy Agency] projects by 2035 that the, the electricity use could double (opens in new tab). And so from electricity use, actually, datacenters use about 1.5% of global electricity (opens in new tab),…
BURGER: Yup.
LUERS: … and that could double, could be between three and fi—even more than double.
But that’s still, from their projections, it would still be less than 1% of global emissions. So even if that would double in that space. So it’s still in terms of a global emissions perspective, which is what the climate cares about, …
BURGER: Right.
LUERS: …it’s a small percentage.
BURGER: Can I just go back for a second, break that down?
LUERS: Yeah.
BURGER: So, so energy is three-quarters (opens in new tab), of … generates three-quarters of emissions. But that includes burning fuel, …
LUERS: Yeah.
BURGER: …transport. And then what fraction of that three-quarters or let’s say just total emissions do we think electricity is?
LUERS: So electricity …
BURGER: Generation …
LUERS: …is about 20% … or, no, electricity … the energy that’s produced is consumed … about 20% of it is consumed as electricity (opens in new tab).
BURGER: Got it.
LUERS: Now, in terms of emissions, about 35% of the emissions from energy is from electricity (opens in new tab) . And part of that is because electricity … the reason there’s that difference is …
BURGER: You’ve got coal plants.
LUERS: You’ve got coal plants, and it’s not as efficient when you do coal plants. You actually get efficiencies when you go right from solar to … in terms of just the energy because you lose a lot of heat in the thermoelectric plants, right? So there’s an efficiency there. But, so about 35% of the energy emissions are from electricity, and electricity production is really the key issue. You know, the key issue of today is, like, electricity and datacenters, right? How are you going to get enough electricity? How are you going to get enough clean electricity?
And that is something that is often more of an infrastructure problem than actually the energy problem. I mean, they’re both true. But it’s getting that electricity in the right location at the right time. And that’s sort of a big …
BURGER: It’s a big messy problem.
LUERS: It’s a big, messy problem that we can unpack a little bit. Because I do think there’s a role for AI, a huge role for AI.
BURGER: Maybe even an optimization problem.
LUERS: Maybe even an optimization problem. Exactly.
MENACHE: Maybe.
BURGER: We’ve got this guy in the room. This is exciting.
LUERS: So we should unpack that. But I think before we go off that, just two points that I think are relevant. One thing is that, which is often not necessarily realized by people who don’t spend their lives … haven’t spent their lives thinking about climate, but to tackle the climate problem, we need massive amounts more of electricity. That is, that’s just part … I said we had 170,000 terawatts of energy. Most of that, to be able to solve the problem, has to come in the form of electricity because that’s what we can decarbonize easiest. So one of those challenges is actually more electricity.
BURGER: Got it. So let me again try to break this down to a simple statement. So we’ve got, you know, about 35% of emissions are due to electricity use …
LUERS: Thirty-five percent of energy emissions.
BURGER: Thirty-five percent of energy emissions, which is three-quarters of the pie. So we can do the multiplication. And, of course, you know, as we decarbonize electricity, there is a probably too slow but ongoing transition towards a lower carbon emissions generation of electricity. You know, you convert a coal plant to a natural gas plant, it gets better. You put it to solar wind, it gets even better. But then the demand for electricity is going up, in part fueled by, you know, the tech industry and building the datacenters and AI, but in part because, like, we have to stop burning fossil fuels.
LUERS: Right. Electric vehicles, turning all of our heating into electric heat—turning everything … the phrase is sort of electrify everything ahead of … the IEA, you know, says we’re in the era of electricity, right?
BURGER: So, so we have a huge pressure on electricity demand. And we need to have it be, you know, low carbon electricity generation, but the demand is just going to go up and up and up with or without, you know, this datacenter buildout and this AI buildout. But, but of course this is happening, and the hope is we can use it to provide a lot of value.
LUERS: Yeah. So there’s one other sort of context which is really important here, and that is that the other concerns that are being raised around this issue of electricity and datacenter growth is at that local community level, right?
BURGER: Right.
LUERS: So datacenters contribute a very small percentage globally in terms of emissions and even in terms of electricity, but they’re really concentrated. They’re one of the most concentrated industries in the world. There’s this great figure if you ever want to look up the recent IEA report on energy and AI (opens in new tab), which really shows the concentration of different industries, and steel’s way on one side, which uses 7% of the emissions, which produces 7% of global emissions. And datacenters are all the way on the other side, which are real in terms of the level of concentration, how close they are together.
And the reason that’s important is that there’s certain pockets of the world where there’s really a lot of datacenters, and they keep going into those areas. It’s changing a bit now because of the dynamics that are happening. But when that happens in the … then in those big … in those areas, yeah, they’re a major user of electricity, right. And when the growth happens quickly in those areas, then it can be …
BURGER: It affects the local grid.
LUERS: … then it can … it can put a strain on local grid. And so there are concerns that are being raised in certain regions in the world about datacenter growth. And, and you know, I’m really optimistic that those are mainly infrastructure problems, and they can be addressed.
And we need to figure out how to do that. And I’m really … that’s why I’m so excited about … in January, we announced our community-first infrastructure initiative (opens in new tab) where we’re really focusing all of our work now on how do we design our datacenter development to ensure that that rapid growth is not a net negative, but actually a net positive for those communities. And that includes, you know, committing to paying all of the prices that it requires to meet our electricity needs so that our datacenters do not drive… prices increase in communities.
BURGER: So I can see how we can get to a net level like, you know, we basically don’t drive up local prices …
LUERS: Yeah.
BURGER: …don’t exacerbate local water supplies and bring it in, however you can do that. But how do you make it a net positive?
LUERS: Well, I think that, you know, we have been saying we’ve been net positive at a global scale, and I think we’re shifting that to say, what do we mean … what does it mean for net positive at a local scale? And I think at a local scale, for example, datacenter water use is, for cooling, can go net positive in the sense that in the datacenters themselves, we are actually beginning to design systems that use essentially zero water for cooling.
BURGER: Right. It’s recycling. Yup.
LUERS: Certainly. So we can, you know, that isn’t in place everywhere, but there’s happening now.
And then we replenish water, so we can do things … so for example, it turns out in many places—water in some cities is lost with leaking pipes, and AI, it turns out, can help identify those leaking pipes (opens in new tab). So even if you get half of that, even if you save half of that, that can be way … our … the emissions from datacenters … the loss … the use of water in datacenters is only a fraction of what we would save. And so you can, you can amplify and save water oftentimes even by using AI, another optimization problem, in many ways.
BURGER: So, so this is more of a, this is more of a commitment from the company to work with the community to get them to a better place. But of course, using our global compute because we’re not going to just run the AI analysis that makes the pipe better in the datacenter that that can be …
LUERS: Oh, no, no.
BURGER: Right. But we’re just trying to … yeah …
LUERS: And it’s not just about … we’re not doing it just with AI.
BURGER: Of course not. Of course not.
LUERS: We’re also investing in training. We’re investing in NGOs. The real focus is to really understand what that looks like. And, you know, my interest is really to say, you know, can we increasingly co-design? What does community positive mean? This is all new because this rapid growth was the first time that this became such a serious issue of concern.
BURGER: Right, right. Well, I’m really glad you’re pushing on this, and I think it’s really important for the communities because if, you know, if we can put enough focus and enough innovation to make these things a net positive, I’ll certainly feel a lot better about it.
Going back to the emissions. So one of the reasons I brought you both here today was because Amy gave a talk at our research showcase, which really inspired me. And you talked about a lot of the emissions being generated through things like transport and inefficient management of some of our large, you know, nondigital systems.
And I … I heard that and I’m like, wow, this might be a place where some of our research could help. And of course, that’s my job, right, is to find places where our researchers can amplify the effect of their work.
So, Ishai, maybe just very briefly, can you talk a little bit about when you worked with the supply chain and you applied your optimizations techniques. And this was even before OptiMind, where you know you could do scenario planning for natural language. What sort of efficiencies were you able to drive just working internally with our business teams?
MENACHE: Yes, so I’ll give you a couple of examples, but starting off with what you mentioned about transport actually. So as part of this intelligent fulfillment service we are actually accounting for shipping costs. And one thing that one can do is account … also take more explicitly into account, you know, emissions and the sustainability considerations when doing the shipping itself, right? So that’s one kind of one option.
BURGER: So we could say, like, get … have the same, like, level of risk and time to delivery, but as your objective just try to drive your emissions down.
MENACHE: Correct.
BURGER: As one example.
MENACHE: Correct. There are other related examples of when, you know, what kind of hardware you use, when you have to fulfill the requirements, and for example, you want to use hardware that has been sitting for a long time in the warehouses. So that also has some implications.
Now, you know, we’ve worked over the years on various systems that, you know, efficiency has been a major goal of us, but sustainability is closely related to efficiency, and I can illustrate it through concrete examples.
So one is virtual machine allocation, which operates at a [very fast] time scale. It’s the process that essentially maps VM requests to physical servers.
BURGER: That’s in our cloud, you know, …
MENACHE: That’s in our cloud. Correct.
BURGER: … when a customer wants to use something. Just again, we’re keeping it, like, not too geeky here.
MENACHE: So one of our goals was to increase packing density, which means that we want to operate the servers close to 100% of utilization. And it is well known, actually, it’s a study by Google (opens in new tab). I guess you’re familiar with it. Actually, you know, as you increase packing density, actually you reduce the power per unit of useful compute, right? So that’s well known. So for example, I don’t remember the exact numbers, but if your server is utilized at 50%, you still consume, like, close to 100% of the power.
BURGER: Yeah, you’ve provisioned it. You know, it’s being transmitted. You know, your chips leak, they have static power dissipation, all the stuff, sort of stuff I used to work on.
MENACHE: That’s actually … Doug knows better than me for sure. You know that’s been his area of research.
So another example that we’ve worked on since 2022 is rack placement. So you have these demands, and you have to decide how to exactly place them, these racks of servers within the datacenters. And there you have to account for power and cooling and, you know, space, and all that. And one of the things that we were able to achieve with our optimization is reduce power fragmentation by 1 to 2%.
So 1% here is huge, you know, and it’s not only about, you know, the cost savings, the money savings for the company, but it’s also about, you know, having to build less datacenters or utilize the datacenters in a more efficient way. And essentially another way to view that is that, you know, if we want to have AI … we want to have AI consumed by everyone in the planet, right? So essentially, you’re making it more kind of … with less, you can do more and have AI … you know, broaden the reach of AI and having it consumed by, you know, eventually all the world.
BURGER: I mean, I think, I think internally in the company and with research and working with our product groups, you’ve been very successful because you’ve, you’ve figured out a bunch of ways in supply chain and in one of our largest businesses to make things run more efficiently, which is great, right.
And, you know, but to the people in the rest of the world, they care, like, you know, is the planet on a sustainable trajectory, right? You know, are their electricity bills going up? And, you know, we’re doing it under competitive pressures to improve our operating margins, which is great for us, …
MENACHE: Right, right.
BURGER: … but we have to think about the world. So, but I think there’s … we have something we can do here.
So now going back to Amy for a sec. In your talk, you talked about, you know, a lot of these emissions and opportunities. Are there a few big buckets you think we can tackle by just being smarter about how we manage complex systems? And what do you think the magnitudes are?
LUERS: Earlier I said I don’t think we can actually solve the climate crisis without AI. And, and I think there are three game-changing capabilities that I see AI really brings to the sustainability process. And I … and one of them is this ability to enable, the way I think about it is, optimization but also understanding and predicting complex systems that we had a hard time doing with or it’s really impossible to do with the traditional analytical tools that we’ve had.
BURGER: Right, right.
LUERS: And the other game-changing capability is the acceleration of discovery, development, and deployment of new climate solutions. And that’s also of course another area that MSR does a lot of work in.
BURGER: Right.
LUERS: And the third game-changing capability that I see AI brings to sustainability is the ability to enhance and augment institutional human and workforce capacity.
And I think all three of these are important and can be game changers if we lean into them. In the context of specifically the managing of optimization or the optimization of complex systems, I think, you know, the biggest challenge that we really have to do is, as I said, transfer … electrify everything, right? And that has many levels of where optimization fits in and accelerating discovery fits in and the third green path. Like, it has various different levels, but one piece is infrastructure.
So the IEA projects (opens in new tab) by 2050, to get net-zero by 2050, which is the global climate goals, at a global scale, we need to more than double the electricity capacity of the world. I mean, now think about that. That’s a, that’s a big undertaking.
And so increasingly, that ability to be able to integrate and manage a system while the climate’s changing and the weather’s becoming much more predictive, that becomes a challenge that if we get it right, we can integrate more. And if we don’t, we’re going to be really going very slowly towards that goal.
MENACHE: Yeah, I want to actually …
BURGER: I saw you twitching in your seat.
MENACHE: Yeah, yeah. So we were talking a lot about optimization, but there’s also the AI itself that helps along the way to make systems more efficient. And, you know, what we do a lot in the team is actually combining AI prediction techniques with optimization, and that is very powerful. And I’ll give a couple of examples. So for example, in the virtual machine allocation set of problems, so what we did, we actually, you know, we predict the lifetime of the VM. Like, how long do we think that the VM will stick in the system?
So there are kind of AI, complementary AI systems that do this prediction for us, and that produces a more … better optimization. When you have, like, intuitively, when you have more knowledge or sort of …
BURGER: Better predictions.
MENACHE: … better predictions. It doesn’t have to be super accurate, but as long as you have, like, decent predictions, you could do much better with packing. So that’s one example. And, you know, in the context of cloud supply chain, predictions, demand predictions, are critical because if you have very bad predictions, then you need to sort of provision ahead. Like, we talked about this, that you have to sort of …
BURGER: You have to buffer capacity, you have to build extra capacity in the system, you’ve got redundancy—it’s expensive.
MENACHE: Right. You don’t know what’s going to happen. So, you know, optimization helps you automate and find the sort of optimal choice. But if your demand predictions are bad, you’re going to have to provision, overprovision these hot buffers. And that also has an environmental effect as well as, you know, cost implications. So I think this kind of blend of AI and optimization, this is something that we are really, I think, really is going to kind of drive things forward.
BURGER: I want to be a little careful here because, you know, we … there’s a lot of excitement about this technology because it’s so disruptive. And, you know, for people listening, you know, I’m an anti-hype guy. Like I used to like to work in areas that were not hot because, you know, communities would get over excited about stuff. But the stuff we’re seeing these advanced models able to do is just jaw-dropping. And, and it is … the capabilities are actually moving really fast.
But those can have bad consequences for society too. I mean, you know, the internet had a set of consequences. Social media had a set of consequences. You know, we are on an unsustainable path right now, trying to get on a sustainable path. And that’s not all tech, that’s just human civilization.
So I really want us to focus, maybe dream a little bit now, like what are the places we can try to steer this tech, you know, AI and all of this compute, to solve some problems that really move the needle for the climate. Because my own hope and goal, and I think, Amy, we probably share it, in fact, I think we all do, is as technologists, you know, steer the technology in a way that helps people in humanity and overcome some of the bad effects because, you know, we’re going to get both as with any new technology.
So like, should we be managing cities with this stuff? Should we build models that just control the grid? Control not in a sense of like, you know, just are able to optimize. Like what are the things we could do five years out, two years out that might be magical and would really move the needle.
LUERS: So …
BURGER: You had your three.
LUERS: Yeah.
BURGER: So maybe we can bucketize it in those.
LUERS: Well, I had three game-changing capabilities. So I would put those game-changing capabilities on problems. Right. So first is, this is what—when I look at AI—are the things that you can put on a lot of climate challenges, sustainability challenges and really make huge differences. And then the question then is, well, what are the three big challenge areas, right?
BURGER: Right.
LUERS: There are a lot of challenge areas. I guess in terms of— we can unpack these—but at the high level, I would say, we’ve already talked about this, but the first one is, you know, enabling electrifying everything, and there are various bullets under that. The second one I would say … I would highlight industrial materials and chemicals, and that’s a discovery-development-deployment type of thing. And it’s packaged like all of these. You can’t just sprinkle AI and it solves the problem, but it can make …
BURGER: It’s real work.
LUERS: It’s real work, right? But I think that there’s a different model in the second bucket of industrial materials and chemicals. I think the thing that I really at least envision that there’s a different way of approaching that because AI exists, and we can unpack that if you like.
And the third one, I would say, is building … enabling the development of a low carbon and resilient secure food system.
BURGER: I see.
LUERS: Food system counts for, you know, about one-third of emissions. And it’s also one of the biggest impacts of climate change itself. One of the issues of vulnerability and food insecurity is a huge issue.
BURGER: Is this fertilizers?
LUERS: It’s fertilizers. It’s also …
BURGER: Water …
LUERS: … like methane in terms of cows, …
BURGER: Methane emissions …
LUERS: … it’s in terms of, you know, and also how we grow food and how we distribute it in terms of supply chains. There’s … food waste is about 8% of emissions (opens in new tab). So there’s a huge amount of inefficiencies. And again, I think the three game-changing capabilities can … there are ways that it can impact each one of those three buckets.
BURGER: Within MSR, we, just in the last month, you know, we’ve had teams design a better X, where X was this percent more efficient using, you know, genetic algorithms or neuroevolution.
You know, these aren’t LLMs, but LLMs, you know, can play a role in this. And I wonder if we were going to go tackle a set of problems, you know, so Ishai’s work could think about taking large, complex systems and figuring out how to run them more effectively. And I think what you’re saying is, could we, you know, could we design a better wind turbine that would generate more electricity, you know, for a given wind speed or just through optimization or …
LUERS: Well, I mean, I think that that would be …
BURGER: … materials, like, for direct air capture, carbon capture. What would be on your list?
LUERS: Yeah. I think in the context of materials, you know, some examples, I mean, of course, cement. Some of the materials, a lot of the issue is tied to electricity. So it’s the optimization of the process. So for steel, for example, you know, a lot of that is what energy you use to do it and how you can optimize that system.
But there are also lots of different materials that can make processes more effective, you know, materials for desalinization, would reduce a huge amount of electricity, and that also connects to this vulnerability and security piece, you know, which I said is a threat.
BURGER: Potable water would be huge if you could do it with low energy.
LUERS: Yeah, yeah, and so the way, you know, one of the things that I have sort of dreamed about is like it couldn’t now … we used to have these challenges, these like grand challenges that, you know, were decades long, that we would have moon shots. And I feel like if we focused AI on different challenges and said, let’s think about those grand challenges in the areas that are really focused where AI can make a difference and say, let’s think of them as like factory moon shots. In other words, like, let’s just say we are going to set up a system to be able to … what are the 10 materials that if we solved and addressed it would really make a difference in energy and food and all the … in society.
BURGER: We need this list.
LUERS: And kind of check them off. You know, like just … and get public-private partnerships to also do that together. And then you have to prepare, though. You have to prepare alongside to be able to have that system so that they can move into society, right.
BURGER: That’s exactly right. And I’m, you know, I was going to pull up my, you know, pull up a paper here. Kristen Severson, who’s at MSR, and three of her colleagues at the University of Washington—I think you know about the work—you know, used a form of machine learning, you know, Markovian processes, to design cement that used … that had algae in it, but is as strong and generates 20% fewer carbon emissions. Now cement is a huge chunk. I don’t remember what it is.
LUERS: It’s about 7%, 6 to 8% (opens in new tab), depending how you measure it.
BURGER: Yeah, so you chop a fifth off of that, and all of a sudden that’s 1 to 2% … between 1 and 2% of emissions. But of course, you know, and it looks like it will be as strong and have all the right properties. But getting it from the paper, you know, in Cell to wide-scale production across the world is a huge undertaking.
LUERS: Totally
BURGER: So, you know, we can solve these problems locally, but then getting them to scale, especially when you’re working in a company that doesn’t produce cement, like how do you do that? That’s a, to your point …
LUERS: Right. That’s why you need it… you need to do this as a public-private partnership. Like this should be thinking about it in terms of, we are going to have a mission to get 10 of these in the next however many years, not trying to just have a moon shot for one in the next several decades, right?
MENACHE: So my list is, like, so number one is everything that Amy said, like …
BURGER: That’s a long number one. [LAUGHTER]
MENACHE: Yeah. So, no, but I should say one of my mentors said like every problem probably has some operation research, OR. You know, when you’re standing in a line, at McDonald’s—that’s where he said it to me—you know, to take this queue or that queue, you know, and like, you know, when you stand in line. So that’s how he was thinking.
But you know, to the point, I think domains that you described, you know, materials, food distribution, electricity, there’s a lot there to, sort of, that we can contribute as optimizers. And in fact, I think we can do a little bit more when it comes to concretely kind of accounting for sustainability metrics. I mean, we talked about that. You know, taking into account more explicitly renewables and such. So I think there is some way to go there.
Number two, I think that, you know, optimization and AI can be used for humanity, for well-being. I’ll give one example. There are all sorts of scheduling systems, and it’s actually a project we started looking into. You know, how do you, sort of, make the system more efficient? You know, you want to reduce cost, but still account, for example, for labor laws and, you know, prioritization on or, you know, union kind of considerations.
And AI can help with that, you know, to actually understand the contracts and understand the fine prints and come up with an algorithm, scheduling algorithms, that are good also for, you know, for the people, right. So it’s, like, you know, for example, if you take into account how many hours drivers, you know, they have this kind of long shifts. How can you sort of prioritize that they have enough rest and account more explicitly for their well-being? So that can be done with a combination of optimization and AI.
Lastly and importantly, I would say, you know, part of the premise and, like, the vision of OptiGuide, which is the project that we’ve been working on in the intersection of gen AI and optimization[1], is making all these kind of complex, you know, tools that we have for making better decisions … and by the way, it’s not only, like, I gave an example of, you know, optimization algorithms, integer linear programming, and all that, but it’s not only that. Think of like very advanced simulation tools that you have also in the electricity space when you model greed and stuff like that. Make them more accessible to end users, to planners, business operators and executives that have to make decisions. So I think that’s also something quite important here that, you know, AI can help facilitate.
BURGER: People, humans get, we get very comfortable with rapid change as long as the change stops. Like we can ingest, you know, if a second moon appeared in the sky, everyone would look at it be like, oh my god, there’s a second moon.
And a week later it’d be like, yeah, there’s two moons. And that’s just how it is, right? And we normalize to it. And so we all think that the society we live in right now is normal, but it is a historical anomaly, you know, exponential population growth, agriculture, industrial revolution allowing a massive increase in population. And our whole economic model is sort of based on exponential population growth fueled by unsustainable fossil fuels and resources. Like, that’s our standard of living. And a big chunk of the world is still really poor.
So I think we have to move to something different, and maybe we just get to this … enough materials to a sustainable future. But to your point, you know, it’s not just about efficiency because these are human beings. These systems we build have to factor in human well-being. And that’s incredibly complex. You can’t do it with, like, a bespoke regulation, you know, regulations hard, policies hard. So my dream is that we can use these very complex systems to kind of evolve and learn that balance. Like, can we actually manage society and its complexity in a way that fulfills the human condition? I don’t know if it’s possible, but it’s, you know, we’re fighting this fight, and it’s really tough.
So I guess for me that’s my hope. I, you know, Amy, you have dedicated a big chunk of your professional life to this topic. Like, what is your aspiration? Like, what would you like to solve? If there’s a problem we could solve in research, what would it be, like kind of what, what would you really like to see happen? What’s your dream?
LUERS: Well, I think to answer that, I’d like to put a twist on your vision of that in terms of … because the way I see it is right now we have built over the years a society that’s based on a broken system. We have, you know, fossil fuels at the bottom, and broken infrastructure and inequities in the world.
BURGER: Massive.
LUERS: Massive inequities in the world. And so right now, anytime we, like put a new thing on that system, it just creates more emissions, draws more problems because we have the same model that we’re doing.
BURGER: Right.
LUERS: The unusual thing about AI and the reason I think it’s so promising and so scary is that it’s the first time we have a tool with such power that can actually change that system that we’re built, we’re based on if we use it to do that.
BURGER: Yes.
LUERS: And so, you know, I think we need to focus it on those things. I’ve written about, you know, five and now it’s become six, but I’ve written about sort of five things that have to happen for us to be able to direct it to change that system.
And the first is, yes, we have to use AI for sustainability, right. I know that sounds sort of trivial, but it’s not trivial because everybody often says, well, there’s a lot of potential out there. And I’m like, yeah, well, the climate crisis, solving the climate crisis is figuring out the potential and then making that a reality.
The second is data infrastructure, and the data and data infrastructure to be to be able to solve these problems.
There’s a lot of … we have to get that data out there right. The third is making access to clean energy and reducing our footprint and supporting communities. Like, those I think are the infrastructure side of it.
BURGER: OK.
LUERS: The next one is what I call governing AI for Earth alignment. And we, with colleagues from around the world, have written… we wrote a paper in Nature Sustainability that sort of outlines this vision of what it means to have Earth alignment in AI in principle. And I think that’s another one. And then the fifth is to skill … upskill the world, like, you know, you said, make it more accessible.
Some of it’s not necessarily learning how to, you know, build agents, it’s being able to, how to understand how do you integrate these tools into your life in a way that actually can drive this change. And I would add one other—and probably is closer to number three than to be in the sixth—but I do think we need to figure out not just [how] to work with communities, but to align the development of datacenters and AI operations with the needs and the trajectory of the electricity grid. And I think that’s not just applying AI to solve solutions through optimization but also thinking about this in an integrated way.
BURGER: Yeah. Ishai, you have a response, and then I have a question for you.
MENACHE: Yeah. So oftentimes decisions are made by multiple parties, right.
So and, you know, each party makes substitutive decisions, and there’s also, you know, they affect what others have to … can decide. Like they change essentially the settings for others that make decisions. So there is kind of dependencies between decisions that I make and decisions that you can make.
And I think that AI can also help sort of can be the glue in this very compound, very complicated systems where there’s, like, distributed decision making. And I think with AI, with all this kind of agentic sort of workflows, we can, you know, take into account moralistic considerations. So obviously in the energy space, as well. So I see a lot of potential in kind of solving for real problems without making too many discounts or, you know, just saying like, you know, I’m focusing only on my, you know, small world. With this kind of combinations of advanced analytics and AI, we can actually get much further with moralistic and more global optimizations that are good for society.
BURGER: How do we take all of that and putting all of these decisions and, you know, complex systems and infrastructure and balancing and put it under control of these massive computational structures that learn things that are now too complex for us to understand and retain human agency?
MENACHE: Yeah. I mean, that’s a tough one. I wasn’t expecting this question. [LAUGHS]
BURGER: Oh, my next one’s worse. [LAUGHS]
MENACHE: We’re out of time. [LAUGHS]
You know, I feel that, you know, there’s not a perfect answer here, but I think that explainability of these systems, of these sort of complex systems, is part of what AI can unlock. So you’re getting these outputs, but you know, as a user of AI, you can also ask questions of like, why? And that’s actually something we’ve been really focusing on in cloud supply chain management is like, you know, why are these decisions being made? Can you just explain to me the alternatives? So actually, with AI, you can explore alternatives quite fast and, you know, having this natural language kind of interface, and then you can get better understanding why certain decisions have been made.
BURGER: Yeah. Ishai, you know, you’re one of the top experts, you know, in optimizing complex systems. You’ve done phenomenal work, and I’m really proud of the work you’ve done and the impact it’s had. Do you have a dream? Like if your work could scale to solve some problems or do something intellectually beautiful, like, would you like … when you look back on your career and you achieved X, you know, what would that be?
The problems now are so important, and there’s so much that can be done. And you’re right in the center, I think, of what’s possible.
MENACHE: You know, I think the dream is to get into areas which we thought that were impossible, like, a few years ago. So like, no, this is not a place where some of your expertise—that’s not for you. …
BURGER: Yeah.
MENACHE: This is, like, you know, there’s legacy, and there’s, like, you know, it’s just too complicated.The data is not in one place. And all that. So we did a little bit with the supply chain, which is, you know, a big challenge there is data.
So I would say, you know, at least from a technical perspective, you know, using AI optimization, advanced analytics to essentially create a unified decision intelligence platform, where the sky is the limit, where you have maybe multiple decision making, multiple considerations. You don’t even know when you start what all these considerations are.
So like having, like, a more interactive system that optimizes in a continuous form. And, you know, one aspect of that is also, which is I think a big challenge, is all these black swan events, rare events, that, you know, optimizations, it’s hard for them to, you know, account for.
So, you know, how do we kind of incorporate optimization in real time and have it reactive and maybe, you know, with a bit less human in the loop, but the human is still an important part in, at least for the foreseen future, in verifying these outcomes and being confident about them. And I would add to that also a part of it is, like, educating humans that are not, like, maybe do not have the same, you know, level of expertise in certain areas.
That’s another thing that we have to think of. How do we explain to humans that, you know, have something to do with the business? How do we explain to them and make them feel comfortable about what the system underneath with all these AI and all these complexities are doing?
BURGER: I, you know, I was listening to it, and it’s really, it’s a beautiful answer actually. And it’s really compelling. I’m going to try to restate it.
Number one, take any system that you can define, no matter how complex, and drive it to near optimal for some criteria—for emissions, you know, managing an energy grid, for example, or how much renewable capacity to provision, or which problems should I … so that, like, if you can achieve that, right, then humanity has an amazing tool to just make … squeeze all of this waste. You know, we talk about, like, wasted food. That’s a massive target you could go after. So just that capability, any system can be brought to near optimal. We just have to define it and then pick one.
And then the other one I loved is how do you plan for resilience if you can speculate on a few of these black swan events? We know that these things are going to keep happening, and they’re going to get worse.
And we’ve seen the hurricane in Jamaica, right? And we just, you know, and North Carolina, and there’s just … so how do we build societal resilience with these rapid changes? That’s an optimization problem.
LUERS: Well, but that’s an interesting question. I would wonder what both of you think about is that, I mean, I think there’s, there is a world that thinks that optimization and resilience are not necessarily always compatible. And so, you know, there is a world where, you know, and I … that resilience and … redundancy in nature is about resilience. You know, that is part of the resilience of these systems. The complexity …
BURGER: There’s a reason why we have two kidneys. [LAUGHS]
LUERS: Yeah, and so, you know, I think a lot about the resilience of ecological systems and, you know, and there is a lot of concern about that. When you optimize it, are we actually making ourselves less resilient? I’m just curious. I know we’re running out of time, but as this tension came up, I was wondering what you guys have thought about that in terms of, you know, how we can move forward with these tools in that space.
BURGER: Well, I’ll jump in there and then turn it over to Ishai. I mean, having designed, you know, with my team, large scale systems that have to be, you know, resilient, you also don’t want massive replication. So there is an art to figuring out where do you over-provision, and how do you do it intelligently so that you get, you know, given some model of failures and byzantine problems like the right outcome. And so, you know, resilience can be just: I’m going to put a ton of food and water and fuel everywhere in case something happens. But you can also be really smart about it. So it really is an optimization problem.
LUERS: But you need resilience in your criteria for optimization.
BURGER: That’s right. You need to know what you’re predicting for and what you’re trying to solve. And then you optimize. And that’s what I was trying to say to Ishai. You optimize around that.
MENACHE: So my first answer was more like Doug’s, is on the system side. And definitely resilience is being accounted for, you know, durability, and also things …
LUERS: Yeah.
MENACHE: … in certain systems, storage systems and such, things that Doug has worked on for many years. I should say that more on the supply chain side and optimization side, it’s a big topic, resiliency.
LUERS: Yeah.
MENACHE: And, you know, and optimization can capture in a formal way measures of risk, right. So you’re not, like, optimizing … necessarily optimizing for the average outcome, average profit, average cost. But you can take into account explicitly risk measures when you’re optimizing, like, value at risk, and other measures that explicitly account for risk. And there are other ways in optimization that you account explicitly for uncertainty, stochastic optimization, where you try to model the uncertainty in the form of, in a distributional form, robust optimization.
Maybe you don’t have the distributions, but you still have some sort of polytope of possibilities in the world, and you just kind of, and you do it … actually network routing, for example, you’re saying like, I don’t know exactly where I’m going to be, but I’m optimizing in a way that no matter where I am, what the state is going to reveal to me, I’m going to do quite well, right.
So there are ways to account for that again, depending, and what I said about AI is that, you know, again, a supply chain planner, a network operator, they might not be familiar with these formal terms.
LUERS: Right.
MENACHE: So how can we make it more accessible to them? Like they have to define the objective now. They know that they have some risk that they want to account for. You know, how do we sort of help them formulate it in a way that is commensurate with what they actually need?
LUERS: Yeah.
BURGER: Amy, you look like you want to weigh in, and then I’m going to bring us to a close.
LUERS: I guess I was thinking that this is the area … you had asked about human agency and this idea of human in the loop to not having automatic. But there’s also another side of it—of how you could have these systems potentially help not just focus on optimization without thinking about resiliency. You know, …
BURGER: Right, right.
LUERS: … if you embed that in the system, then it might have this two-way influence that could have a net benefit.
BURGER: Hundred percent.
LUERS: So, you know, that’s something I think that’s a really interesting area to pursue.
BURGER: I can tell you both that if I were listening to this podcast after hearing the discussion … and I want to thank you both. This was really fun, and I learned a lot today, which is great.
LUERS: Yeah, I did too.
BURGER: I didn’t have somebody saying polytope on my bingo card this morning when I got up. … I would be really curious and maybe even a little bit antsy to see what is that list of those top problems, priorities, and then from you because you talked about it. And so maybe if at any point you get it, I’ll post it along with this. And then which one does Ishai pick up and start working on?
MENACHE: Yeah, let’s talk. [LAUGHTER]
BURGER: You know, it would be great to have some positive outcome. So if we do get the list, if you do pick up one of these and make progress, you know, I would love to report out to people. And I’m hoping, you know, the ties we made today and the discussion we had lead to something like that, …
MENACHE: Absolutely.
BURGER: … that makes the planet better.
LUERS: Excellent.
MENACHE: Thanks a lot.
LUERS: Thanks so much for having us.
MENACHE: It was lots of fun.
BURGER: This was really great. Thank you both.
MENACHE: Yeah, thank you.
LUERS: Great.
[MUSIC]
STANDARD OUTRO: You’ve been listening to The Shape of Things to Come, a Microsoft Research Podcast. Check out more episodes of the podcast at aka.ms/researchpodcast (opens in new tab) or on YouTube and major podcast platforms.
[MUSIC FADES]
[1] OptiGuide is a framework driven by generative AI that lets users design, interact with, and understand complex optimization algorithms through natural language so users can explore scenarios in real time for enhanced decision-making. OptiMind is the small language model that helps with the design and interaction of those algorithms. Both tools are part of ongoing work by the Machine Learning and Optimization group, led by Menache.
Learn more:
- OptiGuide: GenAI for Supply Chain Optimization
Homepage - Microsoft wins Franz Edelman Award for transforming its cloud supply chain (opens in new tab)
Microsoft Signal Blog | April 2026 - OptiMind: A small language model with optimization expertise
Microsoft Research Blog | January 2026 - How Generative AI Improves Supply Chain Management (opens in new tab)
Article | January-February 2025 - Building Community-First AI Infrastructure (opens in new tab)
Microsoft On the Issues Blog | January 2026 - Net zero needs AI — five actions to realize its promise
Publication | August 2025 - The Earth alignment principle for artificial intelligence
Publication | March 2025 - AI transformations for sustainability (opens in new tab)
Microsoft On the Issues Blog | January 2025 - AI tool uses sound to pinpoint leaky pipes, saving precious drinking water (opens in new tab)
Microsoft Source Blog | September 2024 - Sustainable by design: Advancing the sustainability of AI (opens in new tab)
Official Microsoft Blog | April 2024 - Will AI accelerate or delay the race to net-zero emissions?
Publication | April 2024
The post Can we AI our way to a more sustainable world? appeared first on Microsoft Research.
Pennsylvania, USA
This is part of a series of blogs and interviews conducted with our Microsoft Deputy CISOs, in which we surface a number of mission-critical security recommendations and best practices that businesses can enact right now and derive real meaningful benefits from. In this article, Ilya Grebnov, Deputy CISO for Microsoft Dynamics 365 and Power Platform at Microsoft dives into cyberattacks of opportunity and how to prevent them.
When your infrastructure powers thousands of organizations and millions of users, security is not a feature. It is the foundation you build everything else upon. I’m the Deputy CISO for Microsoft Dynamics 365 and Microsoft Power Platform. You may know Dynamics 365 as a cloud-based suite of intelligent business applications that unify customer relationship management (CRM) and enterprise resource planning (ERP) capabilities to help organizations manage sales, customer service, finance, supply chain, and operations more effectively. Power Platform is a low-code suite of tools that empowers both technical and non-technical users to analyze data, build custom applications, automate workflows, and create intelligent virtual agents. It does this by connecting to various data sources through Microsoft Dataverse and integrating seamlessly with not only Dynamics 365 but Microsoft 365 as well.
What might be a little less obvious is that together, these two suites make up what is quite possibly the largest internal business group fully running on Azure at Microsoft. With such a large cloud footprint of our own, and as an important part of the broader Microsoft cloud offering, it’s highly important that we take our digital security seriously. We must remain vigilant against all manner of threats and align our defenses with Secure Future Initiative (SFI) and One Microsoft principles.
I could talk for quite some time about many aspects of security, but I want to focus in on a topic I see mentioned less often than it should: avoiding attacks of opportunity. These are attacks launched by individuals who find ways into systems adjacent to our domains and who move laterally into our space. Maybe they’re looking for our data itself, or maybe they want to use our space as a means locate the company’s crown jewels elsewhere.
To start with, I’d like to cover credential elimination, endpoint reduction, and identity controls. These are strong security practices that everyone can pick up right away. After that, I want to cover the benefits of platform engineering, which delivers some very important security advantages to organizations ready to take it on.
Credential elimination and the benefits of managed identities
Most attackers don’t break into your network. They log in with stolen credentials. While good password hygiene helps reduce this behavior, a more reliable solution is removing credentials from the system entirely. Internally, we rely on a simple principle: if a workload can authenticate without a secret, it should. In following this principle, we have redesigned standards, retired legacy patterns, and eliminated large classes of passwords, client secrets, and API keys across our environment. The fewer credentials that exist, the fewer there are to phish, guess, reuse, or leak.
In practice, credential elimination is predominantly a design choice. Workloads prove who they are without a shared secret. On Azure, the primary mechanisms we use to accomplish this are managed identities (workload identities issued by Microsoft Entra ID) and federated identity patterns that mint tokens just-in-time, with just-enough-access for a specific resource or scope. There’s nothing to store, rotate, accidentally commit to a repo, or forget to expire—which removes a significant portion of potential incident root causes tied to leaked or stale secrets.
Because so many organizations build on our platforms, eliminating secrets in our own infrastructure is just the beginning. We have lent significant focus to making credential-free patterns available end-to-end for customers too. Power Platform Managed Identity (PPMI) gives Power Platform components like Dataverse plugins and Power Automate a tenant-owned identity that authenticates to Azure resources using federated credentials instead of embedded passwords or client secrets. This reduces outages from expired secrets and unblocks makers who previously needed app registrations they didn’t have permission to create. And Microsoft Entra Agent ID treats AI agents, like those created in Copilot Studio, as first class identities so they can be inventoried, governed, and bound to a human sponsor for accountability.
Credential elimination pairs naturally with endpoint elimination, which is the process of reducing or removing public, inbound-reachable endpoints wherever possible. In Azure, when workloads authenticate using managed identities and call out to services, you can:
- Front your data plane with private endpoints/private link, keeping services off the public internet.
- Disable inbound administrative ports (RDP/SSH) in favor of brokered access like just-in-time, bastion, or serial console, and rely on service-to-service OAuth instead of IP-based allowlists.
- Enforce least privileged access at the token level, minimizing the blast radius of misused tokens.
Together, these tactics result in fewer places for an opportunistic attacker to reach. By replacing secrets with managed identities and collapsing public surfaces, you remove the easiest paths in. There are no passwords to stuff, no shared API keys to reuse, and far fewer public surfaces to probe. Even when an attacker gets a foothold nearby, lateral movement is harder because there’s nothing reusable to log in with, and every agent/workload has a distinct, auditable identity you can shut down in seconds.
Platform engineering for security
Opportunistic attackers thrive on inconsistency. Every exception we grant, whether it’s “this team is special,” or “that pattern is fine just this once,” spawns a snowflake architecture with unique configs, unique libraries, and unique failure modes. At small scale, those choices feel harmless. At organizational scale, they can very likely multiply risk and slow incident response. To win long term, we make opinionated decisions centrally and remove room for interpretation, transforming “do the right thing” from advice to policy.
In practice, that means standardizing on common compute and communications, disallowing brittle patterns, and enforcing the same controls everywhere. When we do that, there are fewer places for misconfigurations to hide, and far fewer opportunities for a nearby compromise to pivot laterally into our environment.
Platform engineering delivers the most benefit at scale. I would estimate the most opportune time to take it on is when you reach 500 engineers. Start too early and you risk dampening healthy experimentation; start too late and migration, coordination, and cleanup get exponentially harder. The inflection point is when the cost of fragmentation overtakes the creative benefits of local team autonomy. That’s the moment to set paved paths, publish the guardrails, and commit to consistency.
What to line up before you flip the switch:
- Paved paths worth choosing, including secure-by-default runtimes, libraries, and pipelines.
- Policy-as-code that blocks deprecated patterns and enforces identity-based auth and networking.
- Executive sponsorship to hold the line on exceptions and keep platform friction low enough that teams see the benefits of using it.
Once you’ve decided the time is right for platform engineering, the next question is “who shapes the platform and how do we make trade‑offs?” This is where security and architecture step in—not as blockers, but as partners in defining the paved paths.
Broadly speaking, product and feature teams tend to optimize for success. They want to add capabilities, integrate faster, and ship value. Platform engineering and security, on the other hand, largely focus on minimizing risk. They want to reduce dependencies, question complexity, and enforce patterns that scale safely. Here’s the important part: neither side is wrong. They’re just solving different problems. The key is finding a deliberate balance that satisfies everyone’s needs. You need enough flexibility for innovation within the right number of guardrails to prevent fragmentation and reduce your attack surface.
This mindset shift is critical because it reframes security from “the team that says no” to “the team that designs the defaults everyone can trust.” When security and platform engineering work together, the result is a foundation where controls are baked in rather than bolted on, and one where every service inherits the same protections by design.
I’ll use my own team’s process as an illustrative example of the process. We standardized compute through “core services,” the backbone our application teams use for execution and communications. The tradeoff is intentional: we get a bit less local flexibility for a lot more global safety and speed. When a new defense is needed, one team lands it in core services and over 450 services inherit it, without the need for a service-by-service campaign. That saves time, reduces duplication, and because the evidence is centralized at the platform level, it’s easier for us to both approve changes and demonstrate compliance. We applied the same approach to partitioning and disallowed patterns, to a common communication library (uniform auth, mTLS, retries, telemetry, and policy hooks), and to centralized resource management and telemetry.
Resilience, consistency, and fewer weak links
Credential elimination and platform engineering aren’t quick wins. They’re foundational moves that reshape how you can defend at scale. They demand long term coordination, but the payoff is resilience, consistency, and a dramatically smaller attack surface. Microsoft continues to innovate in this space as well.
Concerning identity, we have delivered PPMI so organizations can securely access their Azure resources without juggling secrets or certificates—and they can use either bring-your-own or platform-managed identities. Next, we’re moving to platform provisioned identities, which are automatically created for each service, partitioned at the cell level, and scoped to the minimum privileges the service needs. Together, these steps materially reduce blast radius if an attacker gains a foothold.
Standardization is the force multiplier for platform security. Because our core services enforce consistent patterns, one change in the platform can protect all of our 450+ services. This saves time, reduces duplication, makes it easier to approve changes, and helps us demonstrate compliance because evidence is centralized at the platform level. This same uniformity is also enabling agent driven automation to help services meet SFI goals at scale—work that would be impractical in a fragmented environment.
Underpinning all of this is the idea of paved paths: opinionated defaults that make the secure choice the easy choice. That’s how we turn security from a checklist into an enabler, and how we make attacks of opportunity far harder to pull off.
Microsoft
Deputy CISOs
To hear more from Microsoft Deputy CISOs, check out the OCISO blog series.
To stay on top of important security industry updates, explore resources specifically designed for CISOs, and learn best practices for improving your organization’s security posture, join the Microsoft CISO Digest distribution list.
To learn more about Microsoft Security solutions, visit our website. Bookmark the Security blog to keep up with our expert coverage on security matters. Also, follow us on LinkedIn (Microsoft Security) and X (@MSFTSecurity) for the latest news and updates on cybersecurity.
The post Making opportunistic cyberattacks harder by design appeared first on Microsoft Security Blog.
Pennsylvania, USA
Guest:
- Jeanette Manfra, VP, Head of Risk and Compliance, Google Cloud
Topics:
- How does "outsourcing" security to the cloud change the intensity of the security vs. privacy struggle for a CISO?
- Does the centralization of cloud make it a bigger target for regulators, or is there a dimension we're missing?
- Does the Shared Responsibility Model actually survive contact with regulators, and how does AI complicate that boundary?
- Can AI actually automate the translation of fragmented rules into evidence, or are we just dreaming?
- How do we navigate the collision between transparency (logging everything) and privacy (recording nothing)?
- What is your one piece of practical advice for leaders helping their teams adopt AI?
Resources:
- Video version
- EP14 Making Compliance Cloud-native
- EP161 Cloud Compliance: A Lawyer - Turned Technologist! - Perspective on Navigating the Cloud
- EP258 Why Your Security Strategy Needs an Immune System, Not a Fortress with Royal Hansen
- EP126 What is Policy as Code and How Can It Help You Secure Your Cloud Environment?
Download audio: https://traffic.libsyn.com/secure/cloudsecuritypodcast/EP273_not266_CloudSecPodcast.mp3?dest-id=2641814
Pennsylvania, USA
Next Page of Stories