
Post training Foundation models RL, Agentic AI() @context_ai
Berlin, Germany
Joined September 2014
- Tweets2K
- Following2.4K
- Followers294
- Likes4K
What do diffusion models, representation learning, reinforcement learning, and test-time reasoning have in common? Each can be seen as learning an energy function or using one to guide sampling and search.
Excited to share our tutorial on energy-based models with @daibond_alpha!
1/5 EBMs are usually taught as one model family among many. In our new tutorial with @du_yilun, we use them as a single lens on machine learning.
"An Energy-Based Perspective on Machine Learning" arxiv.org/abs/2611.14383
Anup Patil retweeted
We breathed life into LLM system.
Introducing Evolving Infra⚙️: an inference framework that adapts diverse and dynamic agent workload purely online and retires fixed PD disaggregation.
vs. NVIDIA Dynamo on up to 256 GB300 GPUs:
➡️ Up to 4.1× lower latency
➡️ Up to 14× lower P95 TTFT
➡️Up to 2.36× more revenue
on real agent traces(Computer Use, Coding and math agents) and opensource models.
🌐 Website: inference-rsi.com
📄 Tech report: alphaxiv.org/en/abs/2611.123…
💻 Code: github.com/mit-han-lab/Evolv…
Which agent(e.g. Pi/Openhands) do you want us to accelerate next? Tell us below 👇
Anup Patil retweeted
this is like, the best cyber news we've gotten in a long time? things working
NEWS: Wall Street execs and NSA have told Congress that outages were caused by patching vulnerabilities identified by AI models, sources say
AI did not cause outages but found so many issues that the rush to fix them before an attack overwhelmed systems
washingtonsun.com/technology…
Anup Patil retweeted
Security isn’t a nice-to-have, it’s a non-negotiable.
So what the industry tried to warn enterprises about 6 months ago is here. The events of this week so far are not a coincidence. AI is making it easy for hackers to attack multiple customers. All incident responders are busy, unfortunately there is no silver bullet. If you haven't:
- run a multi model testing harness across your infrastructure.
- patch your code
- find virtual patching solutions for open source
When it happens, because it will. Your SOC should be ready to detect and remediate in record time. REVAMP your SOC to a real time data lake with proactive ML detectors across your data, with AI running correlation if you want a chance to come out ahead. Band-aid reactive solutions which work on multiple data lakes is not a plan.
Anup Patil retweeted
AI agents do remarkably well in discovering new ideas and solving hard science problems.
Do we understand how LLMs search their way to a final idea?
How do they recover after getting stuck during ideation?
We study LLM ideation by turning it into a two-agent game of question-answering with binary feedback.
Anup Patil retweeted
Never underestimate people’s ability to get used to things. Bad management, promo-driven engineering, perpetual re-orgs; it’s the tech version of the boiling frog
Anup Patil retweeted
💯💯 I'm personally comfortable saying it precisely because it's disingenuous to let PhD (or undergrad/MS) students be misled by things which stick to the old guard when the closed doors conversations are already moving differently. It's worth pointing out that I don't like the way things are evolving because it's inherently becoming more nepotistic in that access based on who you know or recommends you starts to matter disproportionately more and I would like a way to return to meritocracy (I know people will disagree with me on this last bit)
Anup Patil retweeted
About 90% of RNA molecules in our cells are tRNAs. To date, we have mostly been blind to their abundance and regulation in single-cell sequencing. Today, I’m excited to share a new manuscript from my MD/PhD work
biorxiv.org/content/10.64898…
Anup Patil retweeted
Over the last few months we optimized the llama.cpp CI to keep up with the fast pace of development
- We now use HF Buckets to store most of our Github Actions cache
- We use HF Jobs to enroll additional CI runners
More info in this blog post:
huggingface.co/blog/ggml-org…
Anup Patil retweeted
ECB's Isabel Schnabel: 🇩🇪 export-driven business model is "no longer viable in a fragmenting world". Germany's problems are not mainly due to competitiveness, but due to a "sustained weakness in domestic demand that makes 🇩🇪’s economy more prone than others to crises abroad".
Anup Patil retweeted
Totally agree. There are many risks, but AI deciding to kill us all is BS, instead humans will use AI to do bad things (and good things too), just as with other tech (only faster).
absolute fanfic. AI has real, immediate, tangible dangers that we need to focus on, for instance the impact on the job market and making sure random script kiddie can't use it for cyber offense.
But this "oh once it's superintelligent it will take over and ensave us, maybe kill us all!" is pure sci-fi fanfic and distracts from the real (and difficult) issues, which linear probing won't solve.
It's worth repeating the LLMs don't just "decide" to do something and then just "do it" in a void. They generate tokens based on our prompts, only as long as we call `model.forward()`, and on giant clusters with terabytes of weights, that's impossible to "quickly self-replicate".
We also need to regularly remind that the OAI/HF incident was specifically a _guardrail-free_ cyber eval where the model is _instructed_ to hack, to find out how good it is at it.
Some recent personal event made me realize there needs to be more push-back on this doomsday nonsense, and we need to re-focus the discussion on real dangers/issues. I won't argue much in the replies though, not interested in peer-reviewing Terminator fanfic.
Anup Patil retweeted
People ask if I’ve read all the books I own. No, not at all. A private library is not merely a storage facility for once-read books, but a tool box, a spice rack, a medicine cabinet — an investment in future health and happiness.
Anup Patil retweeted
This plot has no y labels and you can tell arbitrarily different stories if you’re willing to manipulate it so it should be entirely ignored.
This slide from the Anthropic IPO roadshow was a big factor in my decision to leave. It’s an insane plot and I’m surprised they made it public.
It shows the compute efficiency of pretraining over time. The y-axis is a **log scale**: they are seeing continued superexponential algorithmic progress.
Mythos was not a one-time deviation, but a point on a superexponential. They could already retrain Mythos with a fraction of the compute originally needed.
It doesn’t strictly imply a software-only intelligence explosion- they have been spending more on R+D. But algorithmic improvements do not require ridiculous creative genius and are on the verge of being automated. You do the math.
Anup Patil retweeted
We'll be seeing a lot more of this.
@katiebindley's diagnosis experience
gift link
wsj.com/health/wellness/how-…
Anup Patil retweeted
How much HIIT is enough?
A meta-analysis suggests ~11 min/session achieves 80% of the modelled maximal VO₂max improvement, vs ~52 min for moderate continuous exercise.
These are estimates, not universal thresholds. HIIT appears more time-efficient.
link.springer.com/article/10…
Anup Patil retweeted
High-quality data is limiting modern pretraining. But it turns out you can reach the same performance with far less data, in 4 lines of code.
Introducing 𝐋𝐚𝐭𝐞𝐧𝐭 𝐅𝐞𝐞𝐝𝐛𝐚𝐜𝐤 𝐈𝐬 𝐚 𝐂𝐨𝐦𝐩𝐮𝐭𝐞 𝐚𝐧𝐝 𝐃𝐚𝐭𝐚 𝐀𝐦𝐩𝐥𝐢𝐟𝐢𝐞𝐫.
Feed a Transformer's own top-layer states back in as input, and every token and every FLOP counts for more. The more data each parameter sees, the more you save.
TL;DR:
- On the same data, latent feedback reaches the same loss with 15% 𝐥𝐞𝐬𝐬 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐜𝐨𝐦𝐩𝐮𝐭𝐞 at 100 unique tokens per parameter, and 46% less at 400 (2-epoch budget)
- At the same number of epochs, it reaches the same loss with 10% 𝐥𝐞𝐬𝐬 𝐮𝐧𝐢𝐪𝐮𝐞 𝐝𝐚𝐭𝐚 at 60M parameters, and 20% less at 778M (1 epoch, 50 unique tokens per parameter)
- Downstream (109 tasks in 37 suites): ahead in all 17 settings at a 4-epoch budget
- Our fitted scaling law predicts the savings keep growing with model size. On a 2-epoch budget, the compute saving goes from 15% at 271M parameters (measured) to 34% at 10B and 45% at 100B (predicted)
🧵 𝐏𝐚𝐩𝐞𝐫, 𝐜𝐨𝐝𝐞 𝐚𝐧𝐝 𝐜𝐡𝐞𝐜𝐤𝐩𝐨𝐢𝐧𝐭𝐬 𝐚𝐭 𝐭𝐡𝐞 𝐞𝐧𝐝
Anup Patil retweeted
Replying to @ClementDelangue
Anup Patil retweeted
This whole thing sucks, and I wish none of it had happened. It’s a really sad day!
But I want to make one thing clear: we’ll keep fighting for open-source AI, harder than ever. This is too important to slow down now. Let’s go!






































