Bryan

Co-founder & CTO @ HUVY, a platform dedicated to accelerate dermatology access using AI. (CE IIB)

[alphaXiv @askalphaxiv]

Yann LeCun and his team can't stop cooking

"LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels"

One of the biggest bottlenecks of JEPA is they are hard to train, and this new research changes that.

They propose LeWorldModel, which shows that a small model can learn a usable world model directly from raw pixels end-to-end.

Sitting at 15M parameters, they made it without needing heuristics and avoiding anti-collapse hacks while staying competitive and planning up to 48x faster.

Making JEPA based modeling much more accessible, cheaper, and stabler.

[Andrej Karpathy @karpathy]

It is hard to communicate how much programming has changed due to AI in the last 2 months: not gradually and over time in the "progress as usual" way, but specifically this last December. There are a number of asterisks but imo coding agents basically didn’t work before December and basically work since - the models have significantly higher quality, long-term coherence and tenacity and they can power through large and long tasks, well past enough that it is extremely disruptive to the default programming workflow.

Just to give an example, over the weekend I was building a local video analysis dashboard for the cameras of my home so I wrote: “Here is the local IP and username/password of my DGX Spark. Log in, set up ssh keys, set up vLLM, download and bench Qwen3-VL, set up a server endpoint to inference videos, a basic web ui dashboard, test everything, set it up with systemd, record memory notes for yourself and write up a markdown report for me”. The agent went off for ~30 minutes, ran into multiple issues, researched solutions online, resolved them one by one, wrote the code, tested it, debugged it, set up the services, and came back with the report and it was just done. I didn’t touch anything. All of this could easily have been a weekend project just 3 months ago but today it’s something you kick off and forget about for 30 minutes.

As a result, programming is becoming unrecognizable. You’re not typing computer code into an editor like the way things were since computers were invented, that era is over. You're spinning up AI agents, giving them tasks *in English* and managing and reviewing their work in parallel. The biggest prize is in figuring out how you can keep ascending the layers of abstraction to set up long-running orchestrator Claws with all of the right tools, memory and instructions that productively manage multiple parallel Code instances for you. The leverage achievable via top tier "agentic engineering" feels very high right now.

It’s not perfect, it needs high-level direction, judgement, taste, oversight, iteration and hints and ideas. It works a lot better in some scenarios than others (e.g. especially for tasks that are well-specified and where you can verify/test functionality). The key is to build intuition to decompose the task just right to hand off the parts that work and help out around the edges. But imo, this is nowhere near "business as usual" time in software.

[Robert Youssef @rryssf]

RIP fine-tuning ☠️

This new Stanford paper just killed it.

It’s called 'Agentic Context Engineering (ACE)' and it proves you can make models smarter without touching a single weight.

Instead of retraining, ACE evolves the context itself.

The model writes, reflects, and edits its own prompt over and over until it becomes a self-improving system.

Think of it like the model keeping a growing notebook of what works.
Each failure becomes a strategy. Each success becomes a rule.

The results are absurd:

+10.6% better than GPT-4–powered agents on AppWorld.
+8.6% on finance reasoning.
86.9% lower cost and latency.
No labels. Just feedback.

Everyone’s been obsessed with “short, clean” prompts.

ACE flips that. It builds long, detailed evolving playbooks that never forget. And it works because LLMs don’t want simplicity, they want *context density.

If this scales, the next generation of AI won’t be “fine-tuned.”
It’ll be self-tuned.

We’re entering the era of living prompts.

[Andrew Ng @AndrewYNg]

Last week, China barred its major tech companies from buying Nvidia chips. This move received only modest attention in the media, but has implications beyond what’s widely appreciated. Specifically, it signals that China has progressed sufficiently in semiconductors to break away from dependence on advanced chips designed in the U.S., the vast majority of which are manufactured in Taiwan. It also highlights the U.S. vulnerability to possible disruptions in Taiwan at a moment when China is becoming less vulnerable.

After the U.S. started restricting AI chip sales to China, China dramatically ramped up its semiconductor research and investment to move toward self-sufficiency. These efforts are starting to bear fruit, and China’s willingness to cut off Nvidia is a strong sign of its faith in its domestic capabilities. For example, the new DeepSeek-R1-Safe model was trained on 1000 Huawei Ascend chips. While individual Ascend chips are significantly less powerful than individual Nvidia or AMD chips, Huawei’s system-level design approach to orchestrating how a much larger number of chips work together seems to be paying off. For example, Huawei’s CloudMatrix 384 system of 384 chips aims to compete with Nvidia’s GB200, which uses 72 higher-capability chips.

Today, U.S. access to advanced semiconductors is heavily dependent on Taiwan’s TSMC, which manufactures the vast majority of the most advanced chips. Unfortunately, U.S. efforts to ramp up domestic semiconductor manufacturing have been slow. I am encouraged that one fab at the TSMC Arizona facility is now operating, but issues of workforce training, culture, licensing and permitting, and the supply chain are still being addressed, and there is still a long road ahead for the U.S. facility to be a viable substitute for manufacturing in Taiwan.

If China gains independence from Taiwan manufacturing significantly faster than the U.S., this would leave the U.S. much more vulnerable to possible disruptions in Taiwan, whether through natural disasters or man-made events. If manufacturing in Taiwan is disrupted for any reason and Chinese companies end up accounting for a large fraction of global semiconductor manufacturing capabilities, that would also help China gain tremendous geopolitical influence.

Despite occasional moments of heightened tensions and large-scale military exercises, Taiwan has been mostly peaceful since the 1960s. This peace has helped the people of Taiwan to prosper and allowed AI to make tremendous advances, built on top of chips made by TSMC. I hope we will find a path to maintaining peace for many decades more.

But hope is not a plan. In addition to working to ensure peace, practical work lies ahead to multi-source, build more chip fabs in more nations, and enhance the resilience of the semiconductor supply chain. Dependence on any single manufacturer invites shortages, price spikes, and stalled innovation the moment something goes sideways.

[Original text: https://t.co/5bdEpQcaob ]

[Eric Larchevêque @EricLarch]

Intervention lunaire de @ericcoquerel, qui est PRÉSIDENT de la COMMISSION DES FINANCES (en charge du contrôle du budget de l’État français).

Il démontre une incompréhension totale de l'économie et de la finance des entreprises.

Dramatique pour l'avenir de la France.

[François Chollet @fchollet]

GenAI isn't just a technology; it's an informational pollutant—a pervasive cognitive smog that touches and corrupts every aspect of the Internet. It's not just a productivity tool; it's a kind of digital acid rain, silently eroding the value of all information.

Every image is no longer a glimpse of reality, but a potential vector for synthetic deception. Every article is no longer a unique voice, but a soulless permutation of data, a hollow echo in the digital chamber. This isn't just content creation; it's the flattening of the entire vibrant ecosystem of human expression, transforming a rich tapestry of ideas into a uniform, gray slurry of derivative, algorithmically optimized outputs.

This isn't just innovation; it's the systematic contamination of our data streams, a semantic sludge that clogs the channels of genuine communication and cheapens the value of human thought—leaving us to sift through a digital landfill for a single original idea.

[Daniel Nguyen @daniel_nguyenx]

o3 is incredible

[Yann LeCun @ylecun]

Too often, we think a task is easy because some animal can do it.
But the reality is that the task is fiendishly complex and the animal is much smarter than we think.
Conversely, we think tasks like playing chess, calculating an integral, or producing grammatically correct text are complex because only some humans can do them after years of training.
But it turns out these things aren't that complicated and computers can do them much better than us.

This is why the phrase "Artificial General Intelligence" to designate human-level intelligence makes absolutely no sense.

[Vik Paruchuri @VikParuchuri]

Announcing surya - a multilingual text line detection model for documents. It gives you accurate line-level bboxes and column breaks.

Find it here - https://t.co/DD2HfwIG9i .

[Philipp Singer @ph_singer]

Whenever you are contemplating participating in @kaggle competitions and you might have heard someone say it is too far-fetched from practical data science work, consider this example:

In the recent Science LLM competition participants learned among many other things:

- How to fine-tune LLMs
- How to properly use RAG techniques for augmenting LLMs
- RAG chunking, embedding, similarity search, and other related techniques
- Synthetic data generation for training models
- How to optimize inference code for optimal runtime on limited HW resources.
- How to fit large LLMs in small GPUs. People even managed to run 70B LLama2 on 2xT4.
- And many other things

I can encourage everyone to browse through the solution posts, lots of great insights there: https://t.co/DI2mq9w5Qd

[François Chollet @fchollet]

My interpretation of prompt engineering is this:

1. A LLM is a repository of many (millions) of vector programs mined from human-generated data, learned implicitly as a by-product of language compression. A "vector program" is just a very non-linear function that maps part of the latent space unto itself.

2. When you're prompting, you're fetching one of these programs and running it on an input -- part of your prompt serves as a kind of "program key" (as in database key) and part serves as program argument(s). Like, in "write this paragraph in the style of Shakespeare: {my paragraph}", the part "write this paragraph in the stye of X: Y" is a program key, with arguments X=Shakespeare and Y={my paragraph}.

3. The program fetched by your key may or may not work well for the task at hand. There's no reason why it should be optimal. There are lots of related programs to choose from.

4. Prompt engineering represents a search over many keys in order a find a program that is empirically more accurate for what you're trying to do. It's no different than trying different keywords when searching for a Python library.

5. Everything else is unnecessary anthropomorphism on the part of the prompter. You're not talking to a human who understands language the way you do. Stop pretending you are.

[Google AI @GoogleAI]

Introducing CHITA, an optimization-based approach for pruning pre-trained neural networks at scale. Learn how it leverages advances from several fields and outperforms state-of-the-art pruning methods in terms of scalability and performance tradeoffs → https://t.co/Efokrj0y6W

[Yann LeCun @ylecun]

I-JEPA: Efficient method for Self-Supervised Learning of image features.
No need for data augmentation, just masking.
Joint embedding predictive architecture, not generative.
And it's open source, of course.

Blog: https://t.co/ZuouZgeEMC
Paper: https://t.co/BoHSnELyw8
Code & models: https://t.co/DgS9XiwnMz

[abhishek @abhi1thakur]

Adobe just added generative AI capabilities to Photoshop 🤯

[Yann LeCun @ylecun]

MMS: Massively Multilingual Speech.
- Can do speech2text and text speech in 1100 languages.
- Can recognize 4000 spoken languages.
- Code and models available under the CC-BY-NC 4.0 license.
- half the word error rate of Whisper.

Code+Models: https://t.co/NIGfUZ8KZg
Paper: https://t.co/W15aEWHGIR
Blog: https://t.co/TFKXFtlPwc

[AI at Meta @AIatMeta]

DINOv2 by Meta AI is the first method for training computer vision models that uses self-supervised learning to achieve results matching or exceeding industry standards.

More details and a demo ⬇️

[Alvaro Cintas @dr_cintas]

This paper is going viral.

Why?

Authors were able to find a way to enable the Recurrent Memory Transformer to retain information across up to 2 million tokens 🤯

In simpler words, it can process and remember vast amounts of data, significantly more than before.

Just so you have an idea, GPT-4 handles 32K tokens (~50 pages of documents), while the entire Harry Potter series is ~1.5M tokens.

Imagine the potential implications of this:

• AI could write entire novels, not just blog posts
• It could aid in complex scientific research by analyzing vast quantities of data
• Enhance customer service by retaining years of interaction history
• It could store and recall the entirety of your life experiences!!

[Barsee 🐶 @heyBarsee]

4. Adobe Adds New AI-Powered Text-Based Video Editing Feature to Premiere Pro.

With Text-Based Editing, you can now create a rough cut just like copying and pasting text.

[Barsee 🐶 @heyBarsee]

5. Meta AI open source AI tool that can "Animate Drawings"

This tool comes with a unique dataset of 1,80,000 amateur drawings that were used to train it.

Access: https://t.co/xX59yNHMv1

[Barsee 🐶 @heyBarsee]

6. The latest text-to-image generative AI model from @StabilityAI is available to anyone.

The quality is almost as good as Midjourney.

Try: https://t.co/8vf2Zj9QDc