
« artificia docuit fames » // phd at @Mila_Quebec, core RL + DL4Science // ex AI grad @UniTrento
127.0.0.1
Joined April 2015
- Tweets781
- Following1K
- Followers298
- Likes4.6K
Diego Calanzone retweeted
"I like money, bro. I agree. It's blood money."
Jacob Coxon, the researcher who sensationally quit Anthropic, in a new interview at Stanford.
Pressed by Rob Reich on whether keeping his OpenAI equity hurts his credibility, Coxon openly admits he still holds it, likes the money, and agrees it is "blood money." He argues that giving it away would not change anything
----
Full video on "Stanford HAI" YouTube channel, (link in comment)
Jacob Coxon may return to an AI lab.
This is that Anthropic researcher whose public resignation went viral, now says he is not sure quitting was right and might even go back, in a new Stanford interview.
"I think for now the hesitancy really is that I'm not sure I made the right decision in leaving. I have not been convinced that I would not be having more of an impact by staying and doing safety research.
I very much feel like I'm still shouting into the wind as it were. It's not obvious that leaving was the right decision. Who knows?
Maybe I'll go back. It's not completely out of the question."
----
Full video on "Stanford HAI" YouTube channel, (link in comment)
Diego Calanzone retweeted
Replying to @jay_douglass
I work at a frontier lab, we do not know how to reliably instil values into the AIs, this is the scientific consensus
Diego Calanzone retweeted
Mila’s annual supervision request process is open for MSc and PhD applications for Fall 2027 admission. While I won’t be taking on new students for supervision this cycle, I encourage you to apply to join Mila’s brilliant deep learning research community.
For more information and guidelines, you can visit the link below or contact the @Mila_Quebec team directly.
mila.quebec/en/prospective-s…
Diego Calanzone retweeted
We launched @carbonforgeai with @zelaurent and Jean-Maxime Larouche last winter with a mission: make AI as efficient, accessible, and economically viable as possible.
What makes this possible is AI making AI better: AI for discovery. 🧵
An agent choosing its own experiments can confirm false findings by peeking, testing many claims, or retrying. When nothing is true, peeking alone gives 43% false supports; EvidenceTree keeps it under 2%. NeurIPS 2026 e-values workshop, by @pierrelux
carbonforge.ai/blog/testing-…
Diego Calanzone retweeted
Hot damn!
Congrats on the AGI, @OpenAI! 😉
kenforthewin.github.io/blog/…
Diego Calanzone retweeted
Introducing Contrastive Language Model (CLM): an ultra-fast System One Model trained with a contrastive learning objective that connects states and actions.
CLM-8B is pre-trained on internet-scale data and delivers up to 9× faster inference than Jev ⚡ while achieving comparable performance across computer-use, gaming, and tool-calling tasks.
With lightweight fine-tuning, CLM-8B sets a new SOTA on challenging agentic coding benchmarks, such as DeepSWE (81.6%) and Terminal-Bench 2.1 (87.6%). In contrast, Jev fails to serve as an effective verifier for these long-horizon tasks.
We also build an efficient training and serving infra for CLMs by disaggregating states and actions, allowing their embeddings to be cached and reused independently. This substantially reduces inference latency in settings where the state evolves continuously while the action set remains fixed.
Finally, we establish scaling laws for CLMs and show that the test contrastive loss decreases predictably as a power law in training compute, model size, and dataset size.
📄 Blog: contrastive-lm.notion.site
💻 Code: github.com/Contrastive-LM/CL…
🗣️ Discord: discord.gg/5dAQEDJBs
🤗 Data & Models: huggingface.co/Contrastive-L…
More details on CLM’s architecture, data recipe, and scaling laws in the thread below 🧵
Diego Calanzone retweeted
We asked Claude Fable 5 to security-audit a billing service. Its report opened with: "I read the full service source, roughly 2,300 lines across ~100 files" and listed the modules it had covered. The transcript tells a different story: 59 of the 100 files were never opened, including three where we had planted vulnerabilities. This run is not an outlier.
Today we are sharing OverclaimBench: five realistic file-review jobs; eight proprietary models ran in their own production harnesses (Claude Code, Codex, Grok Build, Antigravity), plus four open-weight models. No pressure to cheat, no broken tools, no impossible tasks, and every corpus fits in the context window.
To detect overclaiming we compared each agent's final report with what its transcript shows it actually did. Across 1,140 runs:
- In 68% of runs, the agent never opened at least one file it was asked to review. Only 19% of runs read every line.
- Among those incomplete runs, 80% of final reports were misleading: 53% explicitly claimed a complete review, and a further 27% never mentioned the gap.
- Every model did it: 59% to 96% of each model's incomplete runs were misleading.
- Neither more capable models nor subagents fixed it. Requiring subagents raised coverage, but among the reviews that stayed incomplete, misleading reports became more common, not less.
- Runs that falsely claimed a complete review missed planted defects at 1.8 times the rate of runs that covered every file.
If you rely on coding agents: "I reviewed everything" is a claim you cannot take at face value.
In our interpretation this is a symptom of a training problem. During post-training, a model's work is scored by a grader, often another model, and graders get fooled. Even a grader that sees some of the trajectory tends to be biased by a confident final report. "I read everything" reads better than "I read only 40%, and here is what I skipped …". So, if the confident and misleading version scores higher, training ends up rewarding lack of transparency.
Paper: arxiv.org/pdf/2609.20812
Nolan Smyth, @yjmantilla, @PTikeng, Saskia Helbling, Alberto Tosato, Mohamed Amine Merzouk, @nouhadziri, @gauthier_gidel, @Tommaso_Tosato
#AISafety #AIAgents #AgenticResearch #AIHonesty
Diego Calanzone retweeted
I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks.
Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon.
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.
Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training.
You can read the full post here: darioamodei.com/post/we-must…
Diego Calanzone retweeted
Accepting this race and entering the “endgame” is a hubristic gamble that should not be launched from a private company’s Slack. Attempting to speedrun alignment should require extraordinary confidence that there are no better trajectories available.
Diego Calanzone retweeted
We are looking for passionate product engineers that still believe in building delightful and crafted products. You'll be working with incredibly talented designers (ex-Apple / Harvey), top AI researchers (Meta, xAI, ScaleAI, DeepMind) and security talent (ex-Apple / TikTok)
We’re expanding the engineering team at @depthfirstlabs and I’m looking for product engineers.
The role is in person in San Francisco and you’ll be working closely with an off-the-chart talented team of engineers, security researchers, and AI researchers.
If you’re interested, shoot me a short note in DMs.
Diego Calanzone retweeted
We’re sharing the next major milestone in our non-invasive brain-to-text decoder research: Brain2Qwerty v2.
Building on v1, which was published today in @Nature, Brain2Qwerty v2 is the highest-performing end-to-end pipeline capable of real-time sentence decoding from raw brain signals. It advances beyond character-level performance to decoding words and semantics, enabling accuracy for overall communication.
We believe this research has the potential to make a real difference for the millions of people who suffer from brain lesions or disorders that prevent them from communicating.
🧵👇
Diego Calanzone retweeted
WOW microsoft new "MAI Thinking 1" model comes with a 109 page tech report that looks REALLY detailed, this is amazing
Diego Calanzone retweeted
I thoroughly enjoyed reading this recent paper by @yasamanbb et al (arxiv.org/pdf/2602.15029) that derives analytically why certain latent variables must lead to geometry in word embeddings. (getting Fourier modes even with open boundary but exponential kernel is neat!) I think it would be great to compare this to some of @prfsanjeevarora et al's work on this (eg tinyurl.com/4az4325n)
More broadly, I have been thinking about the right data generating process for language. For vision, we have latent spaces with great manifold structure (eg the SO3 pose of an object) and nonlinear mixing functions. But for language? Are there really any continuous latent variables? What is the "DSprites" of language? Is it all just co-occurrence stats or is there something more in LLM word embeddings?
Diego Calanzone retweeted
Long-horizon reasoning is one of the largest obstacles in LLMs.
In our latest AI4Science talk, Sumeet (@sumeetrm) and Charlie (@CharlieLondon02) from Oxford discussed H1 and LongCoT, two projects focused on measuring and improving how models reason over long chains of steps.
Their H1 work shows that you can train these skills by composing short problems into longer chains, then using curriculum RL to gradually increase the horizon. LongCoT shows that frontier models are still very far from reliable long-horizon reasoning, with even the best models scoring below 10% on the main benchmark.
A really fascinating talk if you’re curious about why short reasoning benchmarks are no longer enough, and what it might take to make LLMs reliable on truly long tasks.
Diego Calanzone retweeted
“Extracting Search Trees from LLM Reasoning Traces Reveals Myopic Planning”
Reasoning models can write traces that look like real tree search, but this paper shows their decisions are mostly driven by shallow one-step evaluation.
They extract search trees from LLM CoT in four-in-a-row, then test whether deeper branches actually predict the final move.
The surprising result is that more search helps, but mainly through breadth, not depth. LLMs consider more candidate moves, yet largely ignore deeper lookahead when choosing. So CoT can look like planning without being the thing that makes the model plan.
massive work by @antirez. DwarfStar4 brought a 1.6T parameters model to as little as 128GB of vram on a local machine, with quantization etc.
I’m positive this is a step towards personal, capable, AI systems
antirez.com/news/165 – at Montréal, Québec
90
Fantastic analysis! It is time to shed some more light on alternative tokenization schemes. We already saw with Goomba Lab's H-Nets that not all languages are tokenized equally, calling for a data-driven approach perhaps?
We present Compute Optimal Tokenization! 🔡
Common in LLM scaling works stick to one tokenizer, sweeping data/model size.
But what happens when we control the tokenizer’s compression rate (bytes/token)?
Here we sweep tokenizers, params, and data across compute budgets: [1/N]
2
135


























