What a mess of a week around tech security. I’m counting at least a dozen vulnerabilities or leaks across every layer of the stack: Linux, source control, runtime, PaaS, even certificate authorities. Stay safe out there.
[blog] 30 Months on the Exponential (in Charts). What a wild couple of years so far, and we’re just getting started. Ethan’s got a new book out, and talks about this moment.
[blog] Introducing the Gemini API CLI. This new experimental CLI is a no-dependency binary that your favorite tools can use to interact with the Gemini API.
The next competitive advantage in enterprise AI has less to do with which model you choose than with what you build around it.
That’s an important shift. For the past several years, much of the AI conversation has centered on model capability: bigger models, better reasoning, larger context windows, higher benchmark scores. Those advances have been extraordinary. But as AI moves from experimentation to production, the question facing enterprise leaders has changed—it’s no longer, “what can the model do?”—it’s now, “Can we deliver that intelligence reliably, economically, and at scale?”
Increasingly, the answer depends on the serving harness—the engineering system surrounding the model, including the architecture, memory management, scheduling, and optimization decisions that determine how efficiently model intelligence becomes a production capability.
For technical executives, that changes where AI investment should go—and what they should demand from the architecture supporting it.
Start with the workload, not the model
There is no universally “best” model or AI architecture. There is only the right solution for the business problem.
A customer-facing assistant optimized for fast responses has fundamentally different requirements from a document-processing pipeline that can run overnight. An internal knowledge application and a high-volume customer service platform may use the same foundation model but require very different serving architectures.
The workload defines the constraints. The constraints should determine the architecture.
That sounds simple, but it runs counter to the way many organizations have approached early AI deployments. During a proof of concept, model accuracy naturally dominates the discussion. Production introduces constraints that are much easier to ignore when only a handful of users are involved: latency, concurrency, memory pressure, GPU utilization, infrastructure cost, and unpredictable traffic.
A model that produces marginally better answers is not necessarily the better production choice if it doubles inference costs or cannot meet the application’s latency requirements. Not every workload needs a Ferrari. Sometimes the smarter engineering decision is the model that delivers exactly the performance the business needs without paying for capability it doesn’t.
The objective should be to deploy the right intelligence at the right cost for the workload. That makes serving architecture more than an engineering concern. It becomes an investment decision: where to spend on compute, where to optimize, and where additional model capability simply isn’t worth the cost.
Engineer the system, not the model
Once AI reaches production, seemingly isolated engineering decisions begin affecting one another. Serving architecture influences memory utilization. Memory management affects concurrency. Concurrency shapes batching and scheduling, and scheduling affects GPU utilization and latency.
Only then can techniques such as quantization, speculative decoding, or optimized attention mechanisms deliver their full value. This interdependence matters because organizations can spend considerably more on infrastructure without solving the constraint limiting performance.
GPU memory is a good example. During inference, the model’s weights remain fixed, but the key-value (KV) cache grows with active requests and context length. As concurrency increases, a system can run short of usable GPU memory even when compute capacity is still available. Adding more compute without understanding that bottleneck may simply make an inefficient system more expensive.
The lesson for technology leaders is broader than memory management: Measure before you optimize. Identify the dominant constraint. Address it. Measure the result. Then determine if another bottleneck emerged. Repeat.
Don’t think of optimization techniques as Pokémon cards: The goal isn’t to collect them all. The goal is to apply the technique that removes the highest-impact constraint while preserving the application’s business requirements.
For technology leaders, that demands a different governance question. Instead of asking which AI capabilities the organization can deploy, ask which constraints stand between a proven capability and sustainable business value.
Build for change, not today’s benchmark
Perhaps the most important reason to invest in the serving harness is that today’s leading model will probably not be tomorrow’s. Models will improve. Hardware will evolve. New optimization techniques will emerge. The architecture surrounding them should make those changes easier to absorb, not force the enterprise to rebuild its AI platform every time technology advances.
That makes adaptability an increasingly important outcome of AI strategy.
Organizations that build serving systems around workload requirements, clear service-level objectives, and measurable production economics gain the ability to evaluate new models and infrastructure against what the business actually needs. A model change can become an evaluation and configuration decision rather than an architectural reset.
EXL researchers have explored these production challenges in greater depth in the EXL Engineering the Harness series. It examines how serving architecture, memory management, scheduling, and inference optimization work together to make enterprise AI more efficient, scalable, and economically sustainable.
Increasingly, enterprises have access to many of the same models. Sustainable differentiation comes from how effectively an organization turns that intelligence into business outcomes.
The model provides the intelligence. The harness determines whether the enterprise can deliver outcomes reliably, economically, and at scale.
This content was produced by EXL. It was not written by MIT Technology Review’s editorial staff.
Moderna shares plunged as much as 25% after partial clinical trial data leaked for its personalized cancer vaccine developed with Merck, despite the full results showing promising outcomes for melanoma patients. Analysts attribute the selloff largely to inflated investor expectations rather than disappointing clinical results. CNBC reports: The personalized vaccine, known as intismeran autogene, paired with Merck's immunotherapy Keytruda reduced the risk of recurrence or death in certain melanoma patients by 42% compared with Keytruda alone. The data also showed the shot along with Keytruda reduced the risk of cancer spreading to distant parts of the body or death by 38%, compared with just the immunotherapy. The companies say about 82% of patients who received the personalized cancer vaccine alongside Keytruda were alive without their melanoma returning at two years, compared with 70% of those who received Keytruda alone.
[...] Investors see the targeted vaccine as a possible breakthrough treatment, with the potential to be used in other forms of cancer in the future. It could also offer Moderna a critical additional commercial product in its pipeline, following vaccines for Covid-19, the flu and respiratory syncytial virus. "We assert this relative reduction in recurrence/death is clinically meaningful," said BMO Capital Markets analyst Evan Seigerman. "Such a reaction likely reflects inflated expectations into the full data release, rather than a true indictment on the efficacy of the combo therapy."
UBS analysts said "this is overall a strong data set" and gives more confidence in upcoming studies of the vaccine in kidney and bladder cancers. They added the stock reaction may be due more to massive expectations than disappointing data.
Microsoft has been quietly testing a Windows 11 AI app codenamed Project Pantheon, which works like a version of Recall that only saves what you choose. In the build I used, called Context Vault, you can save web pages, files, screenshots, copied text, audio, and video as “memories,” sort them into collections, and ask AI questions about them; the app can even share that context with Claude and Cursor through an MCP server.
Microsoft hasn’t announced Pantheon, and the app didn’t show up in normal Microsoft Store searches. I got it from the Store after the unlisted listing was discovered, without any invite or access from Microsoft, and used it over the summer, including to plan a beach trip. An easy way to think about this new AI app would be to call it a second brain.
Project Pantheon vs Windows Recall
Project Pantheon is Microsoft’s AI app for saving memories on Windows 11
Pantheon has gone by several names. The app I tested is named Context Vault, the platform is called Context IO, and earlier versions used the name Projects, which still appears in Windows 11’s Settings app. Whatever the name, the idea is to help you save useful content, organize it, and come back to it later with AI.
Context Vault opens to a Collections page, with a sidebar for Library, Chat, Token usage, Timeline, and Connections. Anything you save without picking a collection goes to “Uncategorized memories,” which held 11 items on my PC.
Context Vault (Project Pantheon) with my saved memories sorted into collections. Credit: Windows Latest
The app can also create collections by itself, marked with an “Auto” label. One of mine was a sensible “Weekend Plan” built from a single note, but two others were named “None” and “null,” which is a good reminder that you’re looking at an unfinished test app.
Pantheon integrates with the Windows shell, so you don’t have to open the app to save a file or a note. I could add memories from a browser extension, File Explorer, supported apps, the Windows Share menu, Screen Capture, Click to Do and the clipboard, and audio and video content could be included too.
In File Explorer, Context Vault shows itself in the Share menu next to WhatsApp, Outlook and TikTok, so saving a photo takes a couple of clicks.
Context Vault shows up as a target in the Windows 11 Share menu. Credit: Windows Latest
It also registers its App Actions in Settings > Apps > Actions, which is how it shows up in Click to Do. The two actions are “Save image to Projects” and “Save selected text to Projects,” using the app’s older name.
Context Vault’s actions in Settings > Apps > Actions still use the older Projects name. Credit: Windows Latest
But what intrigues me is the MCP server. Model Context Protocol is an open standard that lets AI apps pull data from other apps, and Microsoft has been building it into Windows 11 since 2025. Using Pantheon’s MCP server, apps like Claude and Cursor could use the memories I saved, so the context wasn’t locked inside a Microsoft app, which is a good idea considering Redmond’s reputation with Recall.
Ask Context Vault’s AI about everything you saved
Once you have a few memories, the Chat page lets you ask questions about them. It suggests prompts like “List my collections,” “Summarize a collection,” “What have I been working on?” and “Draft using a collection,” and there’s a “Remember this” button next to the chat box.
The Chat page in Context Vault lets you ask AI about everything you saved. Credit: Windows Latest
In my testing, Pantheon brought related material together without requiring me to remember where each item came from. Instead of searching through several apps and folders, I asked a question about the collection, and it worked surprisingly well.
The version I tested could also use cloud-based AI models, so it didn’t need an NPU. A Token usage page tracks how much you’ve used, which you won’t find in Recall or other on-device Windows AI features.
That said, an app that could be your second brain could be supercharged if it could utilize local AI models and run on-device, especially with powerful SOCs like the NVIDIA RTX Spark.
Project Pantheon vs Recall
Saving memories may sound similar to Recall, but Pantheon required me to choose what was saved instead of capturing everything automatically. Recall takes snapshots of your screen every few seconds and only works on Copilot+ PCs, and its privacy problems made it one of Microsoft’s most criticized features.
Pantheon is also a separate download, while Recall is built into Windows. The manual approach gave me more control, but it also meant more work.
Windows 11 Recall Homepage. Image: Microsoft
I used it to plan a trip to the beach this summer, and my main frustration was having to remember to save everything manually. Ironically, the one thing Recall does well, capturing everything in the background, is what I missed most. Larger collections were also harder to manage because the app lacked proper filtering, although I understand that’s being considered.
Also, the fact that I can give inputs (or memories) manually makes the case for both of these apps or features to coexist. I can’t help but think about Chinese smartphone maker Oppo’s Mind Space or Nothing’s Essential Space.
Microsoft is considering encryption, Windows Hello and IT controls
I understand that Microsoft has considered several additions that could make Context Vault safer and easier to manage if development continues. These include encryption at rest, per-app access controls, Windows Hello protection similar to Recall, user consent controls, IT policies to disable the app, Settings integration, and an API that would let other apps access your memories.
Interacting with text in Recall. Image: Microsoft
Microsoft hasn’t confirmed any of these, and it hasn’t committed to releasing Pantheon. But after Recall, the consent controls are the part I’d want first, especially since Microsoft already had to promise that Windows 11 will ask before AI agents read your files. Microsoft would also need to explain where memories are stored, when cloud services are used and how you can delete your data.
Is Project Pantheon useful?
Well, if you ever wanted to take the load off your brain’s memory and hand it over to a faster second brain, you’ll definitely find Project Pantheon helpful. After using it over the summer, I felt the same. Having one searchable place for material from the web, documents, screenshots, and copied text saved me time, and it felt better suited to focused work than Recall, which records everything whether you need it or not.
Still, Microsoft already has Recall, Click to Do, Copilot, and AI agents on the taskbar, and Pantheon would be one more AI feature asking for your data. If Microsoft does ship it, saving a memory has to be as quick as copying text, or most people will forget to use it, just like I sometimes did.
Also, knowing Microsoft, I would recommend the folks at Redmond be very careful while presenting Project Pantheon, or else Cupertino might come up with the same idea a couple of years later and call it a macOS breakthrough!
Discord's latest update released today has a big improvement for Markdown previews: they should now render correctly instead of with "a bunch of symbols," according to the company. LLMs and AI agents often rely on Markdown files for instructions and context, so these improved previews could be helpful for people sharing information about AI projects on the platform.
"When you upload a Markdown (.md) file, you'll now see the file rendered all nice and neat and pretty instead of with a bunch of symbols," Discord says. "This includes tables, formatting, all that good stuff. Totally great for when you're doing a casual Sunday-night sharing READ …
Log in
