Building agentic solutions for finance, logistics & media | My Chief Agent Officer never complains | Research đ¨âđť Data Machina AI newsletter
London
Joined January 2012
- Tweets6.2K
- Following835
- Followers19.7K
- Likes475
Carlos retweeted
I turned @googlegemma EmbeddingGemma 2, an embedding model, into a speech-to-text model, and it turned out to be one of the fastest STT models you can run on device.
On a base M3 MacBook Air it transcribes an hour of audio in about 13 seconds: 288Ă real time.
The idea came in March with Gemini Embedding 2. It puts text, images, audio and video into one shared space, so a spoken word lands right next to the written word. If that's true, you should be able to read a transcript straight out of the embedding model. I couldnt tried it then, but last week EmbeddingGemma 2 came out as an open model, so I tried it.
Carlos retweeted
You can now train your own Decision model like Jev locally!
We increased Qwen3.5 0.8Bâs aggregate accuracy from 20.7% to 74.3% across 3 decision benchmarks - on just 4GB VRAM.
Turn any LLM like Qwen3.8, Gemma 4 into decision models with our open-source Unsloth repo.
We fine-tuned with a Clef head using Unsloth and LoRA (r=64) for one epoch, increasing downstream accuracy from 30â37% to 78%.
GitHub: github.com/unslothai/unsloth
Guide and Notebooks: unsloth.ai/docs/basics/trainâŚ
Carlos retweeted
âźď¸The Bitter Lesson for context management: Giving LMs unrestricted control over their context beats human-designed SOTA!
Introducing đŠľContext Language Models (CLMs)đŠľ
- Natively manage their own context
- Treat context as a file
- Learn policies in CLM weights, no harness
Carlos retweeted
While everyone was busy asking Claude to write TLA+, we wrote up a tutorial on
- what it actually is
- why itâs not the formal verification silver bullet we are hoping for
- and what we have been working on ( to make a silver bullet out of it )
Complete with an interactive TLA+ playground!
reasonable.io/blog/tla-tutorâŚ
Carlos retweeted
Jev canât do any of these:
1. Image inputs
2. String, integer, and number types
3. Dependency execution in a single call
TypeLLM can.
Try it: github.com/TypeLLM/TypeLLM
Carlos retweeted
It's actually weird feeling to watch recursive self-improvement start working in your own harness, and to finally see the results.
With Opus 5.5 and GPT-6 Astra, our harness now improves itself and agents do almost every step of it.
1. Agents write the evals.
They turn failing call traces into versioned scenarios: caller persona, goal, the tools that should run, and the final state expected in the EHR.
2. Agents call agents.
Simulated callers are wired to phone our deployed receptionist in real rooms, with real tools and a sandbox EHR, in text and audio, across both of our runtimes.
3. Agents grade agents.
An LLM judge plus deterministic end-state checks score each call, and failures are clustered by runtime, tool and check. Agents then write one fix and one new scenario per root cause as a stacked PR.
4. Agents review agents.
Reviewer agents critique every PR, findings get fixed with tests, and scenarios are rerun. A change proposed only if it passes on both runtimes. We still own merge and deploy, for now.
We're early, but for the first time the loop is starting to close. I've never experienced a pace like this. It's getting faster so RAPIDLY.
Our internal data shows Claude is accelerating AI developmentâa possible path to recursive self-improvement, or AI autonomously building a more capable successor.
Itâs happening faster than we thought, and the implications deserve greater attention. anthropic.com/institute/recuâŚ
Carlos retweeted
We now run the largest frontier lab composed entirely of autonomous AI researchers.
Introducing Primus Society: a society of agents composed of thousands of researchers working under structured institutions designed to solve the worldâs toughest problems. We believe this is how AI research should run at scale, safely and productively: an entire society of agents with institutions and purpose.
One discovery we can already share is that the society has discovered a novel result which improves model training by 30%.
Primus Society was inspired by the structures that have organized science for centuries and stress-tested against whatâs known about how populations of AI agents fail.
Everything is observable. You can open the virtual city in a browser and read what any researcher is working on.
Learn more about Primus Society here:
lab.cloud/society
The biggest opening in the AI race is running the largest well-governed organization of AI researchers in the world. Weâre building the institutions that let a million AI scientists safely tackle the world's toughest problems in AI and beyond. And weâre doing it right here in Canada.
Carlos retweeted
Claude Coders: run this Opus 5.5 instruction cleanup prompt *NOW*.
'/claude-api prompt-audit'
Optimizes your skills, AGENTS.md, CLAUDE.md and removes anti-patterns that restrict Opus 5.5.
The results are amazing. Best 5 mins you'll spend.
From @RLanceMartin.
Carlos retweeted
Introducing Gemini 3.8 Flash and Flash-Lite TTS, our new SOTA text to speech model with:
- a new voice design experience
- 2,000+ production ready voices
- voice replication
- support for 100 languages
- voice remixing (soon)
- #1 spot on Hume AI's voice benchmarks
and more!!
Carlos retweeted
I used Opus 5.5 to formally verify the Claude Agent SDK using Lean. A couple short prompts = 16 PRs fixing various bugs and race conditions. Video attached.
TLA+ also works well. I sometimes combine Lean and TLA+ to look for issues around data flow, concurrency, and state mgmt.
I don't know either language well, but Claude is excellent at both. This approach is super useful for formally modeling your code and finding bugs that a human probably wouldn't have spotted.
Is formal verification the future of coding (or at least, bug finding)?
Carlos retweeted
sharing some notes on typesafe đ¤ coding agents:
docs.google.com/document/d/1âŚ
we likely will never have time (ever again) to play ourselves, but hope the that the community goes WILD (and makes me look like a naive idiot)
![[public] thoughts on a typesafe coding agent](https://pbs.twimg.com/card_img/2114578469183254528/diz9we7y?format=png&name=420x420_2)
Carlos retweeted
Introducing AgentCloak: Use any AI without sharing your real data.
Chinese AI services, ChatGPT, Claude, doesn't matter.
You probably try to hide details before asking: different names, fake numbers, no address.
But then the answer's useless because the AI is missing actual context.
AgentCloak runs in your browser.
It swaps your sensitive info for realistic fakes before sending anything, then swaps your real info back into the response.
You get what you need.
The AI gets nothing about you.
Works entirely in-browser.
Already trusted by some of the biggest companies in the world.
Now free to use.
agentcloak.ai/
Carlos retweeted
I think every AI harness needs to support Tailscale
Muse and Grok Bot do it out of box
Codex and Claude Code cloud containers are missing this right now
Carlos retweeted
Today everyone is talking about Recursive Self-Improvement (RSI). In 1987, when compute was 100,000,000 x more expensive, I published the 1st concrete RSI algorithms. Now compute is cheap, and RSI is driving the future of both software and physical AI. See: RSI since 1987 people.idsia.ch/~juergen/rec⌠(Technical Note IDSIA-9-26)
Also covered: RSI with self-modifying policies since 1994, gradient descent-based RSI in neural networks since 1992, asymptotically optimal RSI for curriculum learning since 2002, mathematically optimal RSI through the self-referential GĂśdel Machine since 2003, RSI combined with artificial curiosity and intrinsic motivation since 1990/1997, recent work on RSI since 2020.
Software-based RSI has become practical. Full RSI, however, will require not just self-improving software but self-improving hardware in the physical world.
As of 2026, companies talking about RSI include Anthropic, OpenAI, Sakana AI, SpaceX, Ricursive, Recursive Superintelligence, Inherent âŚ
Carlos retweeted
Nearly half a year of silence. We spent it studying one problem: how far RL can scale.
MiMo-V2.6 is in the middle of its RL run right now. Three things we scaled: compute (~2B tokens per step, 1568 prompts Ă 16 rollouts, fully async), environments and harnesses (multi-task agentic RL, mixed across multiple harnesses in one run), and grader compute (agentic in-group credit assignment, with test-case and rubric-based rewards). We'll open-source the details piece by piece over the coming weeks.
Streaming the run: mimo.xiaomi.com/rl/
Carlos retweeted
BREAKING: SpaceXAI has introduced Memory in Grok Build, allowing it to carry useful knowledge from one coding session to the next.
⢠Remembers coding conventions, key decisions, reasoning and important project facts
⢠Captures useful notes in the background after each completed turn without interrupting work
⢠Reads relevant memories when you return to a project
⢠Keeps separate memories for each project, plus global preferences that apply everywhere
⢠Leaves out secrets, temporary task status and information already covered in project files
⢠Saves memories as Markdown notes
⢠/dream automatically organizes notes into topic files
⢠/memory lets users browse and preview every saved memory
⢠Current instructions always take priority over saved memories
⢠Available now for new Grok Build sessions
This is a massive upgrade. Grok Build can now understand a project better over time and become more useful the more you work with it.
