
Founder Rayo, Cofounder FlipAI; ex-Head of Eng #AmazonComprehend, Creator #AWSDeepRacer, Serverless MapReduce. Tweets on RL, Agents Coding, Sim2Real, Cats
SF
Joined September 2009
- Tweets4.8K
- Following2.6K
- Followers2.1K
- Likes6.6K
Wow Biryani Onigiri. Something tells me this is going to be big in India in a few years
【ビリヤニおむすび、セブン-イレブンでエリア限定発売】
セブン-イレブンさんと一緒に、ビリヤニのおむすびを作りました。
何度も試作と調整を重ね、ようやくビリヤニ大澤らしい味わいに仕上がりました。
バスマティライスのおいしさ、スパイスの香り、ビリヤニならではの軽い食感が出るよう、工程の一つひとつを丁寧に仕上げていただきました。
明日11月17日(火)から、墨田・銀座・船橋・習志野エリアのセブン-イレブンの一部店舗でテスト販売が始まります。
お近くで見かけたら、ぜひ手に取っていただけるとうれしいです。
※店舗によっては取り扱いがない場合があります。
#ビリヤニ大澤 #ビリヤニおむすび #セブンイレブン
2
249
🎉 Launching Rayo @rayohq: Teach AI Agents Your Business
Rayo clones your enterprise workflow, creates a Twin: RL environment where AI agents can explore, learn judgement on your past cases, reinforced by expert feedback. Rayo executes just like humans, hence need zero integrations: No API, No MCP.
Rayo post-trains the agent harness, not the model weights, keeping learned behaviors explicit, versioned, and auditable. This is especially important in regulated industries, where every decision must be traceable.
The real measure of success is not just how effectively agents learn in our environments, but how well that learning translates into accuracy and reliability in production. We've seen this across multiple customers, including @NiumGlobal.
DM me for deploying Rayo or hiring AI Engineers, RL researchers.
rayo.com/
great event, great biryani, great sweets. Thanks @rajoshighosh and team!
Founders, if you didn’t get your 🪔Diwali 🪔 fill over the weekend, come join us tomorrow for what’s turning out to be quite a feast.
We have Hyderabadi, Donne, Malabar and Kolkata Biriyanis (and a veg one too).
What’s a Diwali without an extended feast. RSVP 👇
(Lowkey worried we ordered too much food. So come hungry!)
2
7
294
This will be the new norm. Simulate before production! I can help. DM me
Building an agent isn’t one-and-done.
Before it can solve anything, you need two things: a clear definition of “good” and a clear definition of “done.”
Our paper takes one step in that direction and was just accepted to AAAI 2027 🎉
5
215
This is just for coding, imagine this for rest of your operations!
BREAKING: Revolut moved ~2,000 engineers onto self-hosted GLM for coding and pushed Claude Code and Codex to the exception list.
Per David Abitbol, who leads AI ops there: 56 Nvidia B200s, a blended 30 cents per million tokens, roughly a tenth of what they paid Anthropic and OpenAI. Engineering lead time improved 70%.
Here's the part that matters: the gains showed up when tokens stopped being rationed. Engineers quit saving AI for the hard tickets and pointed it at everything.
Frontier models become the exception, open weights become the default. That's a pricing problem for every lab heading to IPO.
h/t @OliverMolander
Would you move your team off frontier APIs if the bill dropped 90%?
3
238
This just happened today at a conference party, sketchy looking person without badge
Elsewhere: Want some white powder?
SF: You looking for some good data? I got some new stuff.
3
7
303
CAP taught us to design for failure. LLM inference needs the same lesson. Every production LLM system trades latency, availability and behavior. Most make that trade off by accident.
*hot take* Most RL environments today are a data capture wrapped around a unit test. Not true environments that model the real work
8
299
Agentic Sim2real is actually closer to robotics sim2real than you think, based on my experience building a lot of RL SIM environments over the last decade, I capture notes on how it applies to this new age industry.
sunilmallya.github.io/sim2re…
1/ There's a gold rush underway. A wave of companies now sells RL environments to frontier labs, and too many of them are selling the same thing: mocked APIs, synthetic data and a grader, packaged as a "world."
This will get saturated very soon
1
2
158
As much as I dislike how often hill climbing gets thrown around, as an RL guy, I secretly do enjoy it. My mind goes, only if they knew the significance of gradient ascent in RL
3
165
If you are at #COLM2026 and want to chat about RL environments and automated evals, DM me.
I'm trying to fix the env creation, Too many RL environments today teach agents to pass the benchmark and not to operate/generalize in the env.
11
361




















