GPU support in AWS Lambda with Lambda Managed Instances!
I've been waiting for this one for a long time. Custom model inference, speech-to-text, and media processing can now run as Lambda functions triggered by S3, SQS, or EventBridge, with no drivers or clusters to manage.
Thanks to the team for getting this done, along with all the other recent serverless launches.
#AWS #Lambda #GPU #LMI #Serverless
https://t.co/SuMTKUdHip
Thoughts after talking to lots of engineers in London in last couple of days:
• Everyone's moving into the same direction and getting more and more agent-pilled, but all on individual speed. No group of engineers right now is uniformly agent-pilled.
• Becoming agent-pilled is a one-way door. You only realize this once you've gone through. Many think they "get it" but haven't made the switch yet.
• Wild how many orgs are still treating agents as "thing that writes code on my computer" instead of robot-engineer that can help with whole SLDC.
• Teams are drowning in PRs and code review. No process has been changed yet, even though this is the biggest dev productivity lever of last 30 years, engs still work off tickets like nothing happened. Wild.
• Barely anyone realized yet that, hey, even if the model isn't good enough to write the code the way you'd write it [doubt it], they can surely write all of the peripheral tooling to speed you up.
• Very few engineers have imagination to come up with better tool. Instead of reviewing code in diffs, line by line, think: if I could wave a magic wand and make tooling disappear, what tooling would I want to gain confidence in this change being correct? Full instrumented/traced end-to-end runthrough? Demo videos? In-production probes? High-fidelity and anonymized production data fed through test runs? New testing environments with the greatest debugging MCP support one can imagine so agent can inspect everything?
It's time to dream big.
But seems like most people's dreams don't reach further than "this agent works off the tickets assigned to me 10x faster than I can."
this is one of those reasons why I don't think most people will really generate their own software
good software is a set of strong opinions and I'd rather have something that's been really thought through by someone else
posted this into our Rex discord today... the new `𝚛𝚎𝚡.𝚞𝚒` primitives provides a new set of simple but powerful methods to build interactive scripts within Rex that require user interaction.
Now you can build the workflow of your dreams, that little action you always wanted or just a pizza topping selector.
I wrote some docs up that go into detail about how this feature works https://t.co/4UXo1zSrBt
My belief is that automatic model routing is fundamentally unsolvable, because you can’t read the mind of the user.
Users almost never provide enough context (or don’t even know themselves) to determine the scope of what the session needs.
A few people asked me to comment on why Astra/Fable usage seems quite low in this new dataset.
I can't know for sure, but the #1 thing I hear from customers is that Opus/Sol/OSS models have simply gotten "good enough" for a huge fraction of enterprise work, especially when combined with routing.
https://t.co/eNa5ap6FdF
The "whiteboard defense:" I should be able to pull you aside at any moment and ask you to explain any customer-facing system you've shipped. You should be able to clearly explain how it works and defend the decisions you made. This is my benchmark for responsible AI usage.
I don't expect line-level familiarity with the code. I don't care if you remember the exact function name or implementation detail. You may not even know it. I don't care.
But if I ask "why did you do X instead of Y?", "what happens if this actor behaves maliciously?", "what data structure did you use here and why?", or "where does this fail?" you should be able to answer confidently.
For PoCs, demos, experiments, whatever: I don't care. Generate 100% of it and understand none of it. Speed over quality every time in those specific scenarios.
But if you're shipping customer-facing work, you can't be shipping things you don't understand at a high level.
AI interfaces nowadays are still pretty much turn-based.
In my PhD thesis, I thought about what interaction would look like if we could go beyond taking turns.
One example is intra-turn interaction:
hot take: unless you are willing to reimplement git / github (like anthropic), I think monorepos are an antipattern in an age of agents
the alternative is to keep things light and make a lovely cli/mcp for agents to do search over your org's code and docs without a checkout
In 2004, I picked @SwamiSivasubram as my intern. It was probably one of the best decisions I have made as Amazon's CTO. He learned early that you can only test a system properly in a high-fidelity copy of the environment it will actually run in. AWS Emulator, launched today and written and tested entirely by agents, gives you that kind of environment for more than 200 AWS services. Read Swami's guest post on what the engineer's job becomes when agents write the code: https://t.co/yv8zVjkQnE
What if we could manipulate UI at the speed of thought?
Built this voice coffee ordering system with @typesafeai Jev. The drink and any updates show up while I’m still mid-sentence, ~200 ms from word to screen.
Feels pretty magical!
AWS quietly removed a ton of undifferentiated heavy lifting in running secure agent sandboxes yesterday and I don't think people outside my normal circle are aware of how useful this is going to be: https://t.co/6P2eWZd6FG
I gave Fable 5.2 a PDF of one of Leonardo da Vinci’s 500-year-old notebooks and it made his inventions real for me to interact with.
Such an exciting time for education. Anything from history that’s been stuck on a page can now be something you explore.
During pre-release testing, Claude Fable 5.2 and Claude Mythos 5.2 worked on ambitious projects in many different fields. Here are a few.
From a one-paragraph prompt, Fable 5.2 designed three working electronic devices, including this home voice assistant hub.
I had a few spare tokens and Opus 5.5 on medium is capable of incredible things without hardly any usage.
So here's a ~1.5x to *42x* speedup in Rust's "percent-encoding" library which runs on most HTTP requests (anything with params)
Heavily inspired by @yagiznizipli's Ada-url