Yuri Makarov

Meta has recruited 3 researchers from Google DeepMind: Tianhe Yu Cosmo Du Weiyue Wang The trio previously teamed up at Google DeepMind, where they helped build systems that cracked International Mathematical Olympiad (IMO) problems at gold‑medal level. Big things happening for Meta.

5

A couple of years ago, I gave a talk where I suggested the focus on LLMs could be Nvidia’s undoing. I was obviously spectacularly, ridiculously wrong... but part of me wonders if the argument has some merit, but the timing was off… My reasoning: LLMs aren't AI's final form. Token-based autoregression has clear limits. Nvidia was over-optimizing hardware for LLMs. Sigmoid curves look exponential at the start—we're likely in the diminishing returns phase of scaling. Fast forward to now (or last month.. anyway): Yann LeCun yet again calls auto-regressive LLMs "doomed." Each token compounds error probability exponentially. "Not fixable without a major redesign." He's pushing JEPA as an alternative (link in the comments for the curious — cool read). So what? Well, Nvidia (and the industry as a whole) have spent billions optimizing GPUs for transformers. If transformers hit a ceiling and we need a major architectural shift (like CNNs → transformers), current GPUs will devalue faster than depreciation schedules suggest. Nvidia and by extension much of the AI compute ecosystem could take a serious hit. For those of us concerned with AI risks (both practical and financial), this cuts both ways: Upside: There's massive value in building safety systems around current, and future generations of AI models. AI will always have problems, but new architectures will bring new solutions—and entirely new problems. De-risking AI at all levels of the stack will likely continue to be difference between PoC and production. The ride has just started.  Downside: If Nvidia crashes, AI funding tightens up. Unlikely to be an extinction level event, but will likely thin the herd of startups, and potentially concentrate the AI winners to a few large platforms. Either way — I stood up and proposed selling Nvidia in late 2023. Please form an orderly queue for more financial advice. https://lnkd.in/e2cGHPFj

19

3 Comments

The leading Q&A site for developers Stack Overflow has seen question volume drop to levels not seen since before 2010 📉. Coding gives a glimpse into the new realities of information retrieval using generative AI. Source: Marc Gravell, https://lnkd.in/d2Jjxyqx

15

2 Comments

Yann LeCun's new startup is built on a premise that 2-3% of humans might disprove. His thesis: “AI needs multimodal sensory grounding to truly understand the world. LLMs fail because they can't do what we can - e.g. imagine a cube rotating in space.” But in 2015, a condition called aphantasia was coined. People with it have no mental imagery at all. Total darkness when they close their eyes. And yet... physicists, Disney animators, the biologist who led the human genome project, novelists with 50M+ books sold. All are aphantasic. Here's where it gets interesting though. Aphantasics still see. They take in sensory data; they just can't replay it. Their brains encode the world differently. Researchers found they retain "spatial imagery" (navigating a room, counting windows) while lacking "object imagery" (picturing a horse). One described it as closer to echolocation than true sight. They convert experience into abstracted concepts, not sensory recordings. So... what does this actually mean for LeCun's argument? Maybe the question isn't whether AI needs sensory “input”, but whether it needs sensory “replay”. Aphantasics suggest you can build functional world models from abstracted, non-imagistic representations. I'm not saying LeCun is wrong. Text prediction alone probably won't get us to robust AI. But "more modalities = better understanding" feels too simple. The real question: what kind of encoding actually matters? #AI #WorldModels #LLMs #Cognition #Neuroscience

106

48 Comments

Over the last decade, most text-generation systems have been Autoregressive. A newer approach diffusion-based language models is now showing promising early results. Instead of generating text sequentially, they iteratively refine full sequences. This enables bidirectional context, better infill, and more coherent outputs. Recent work (e.g., LLaDA, data-constrained training studies) provides useful empirical comparisons between diffusion LMs and transformers, along with their current limitations. Read full blog : https://lnkd.in/dH3s65Eu Author: Deepak C Nayak

38

2 Comments

LLMs are getting smarter, but they can't stop talking. That chattiness is a very real problem. Verbose responses dilute the user experience, drive up inference costs, and make certain workflows completely impractical. Shortening responses without corrupting meaning or accuracy has a very real impact on cost and user experience. Traditional approaches to evaluating verbosity come with tradeoffs. Metrics like BLEU and ROUGE focus on reference overlap, but most real-world tasks do not have reference answers. LLMs can be used as a judge, but already have a bias towards verbosity. It's a sticky problem, with implications for the viability of certain agent workflows, conversational systems, and real-time interactions where no canonical answer exists. That's what makes this paper so interesting. ConCISE quantifies non-essential content without relying on reference texts, making it especially useful for open-ended tasks (such as agent workflows). Their approach opens the door to a scalable, automated way to evaluate response brevity.

19

1 Comment

The move toward multimodal training makes a lot of sense to me. The idea that general intelligence could emerge from models trained primarily on one type of data (text) always felt like a stretch. The world is beautifully more complex than that.

4

Richard Sutton – Father of RL thinks LLMs are a dead end  Large language models aren't capable of learning on-the-job, so no matter how much we scale, we will still need a new architecture to enable continual learning. https://lnkd.in/g6M3AESB

12

2 Comments

Just put out some new work on arXiv, where we introduce a better recipe for collecting LLM post-training data when using GRPO as the RL algorithm. Collecting expert annotations is expensive and budgets are limited. The key question: which examples are actually worth buying? We find that focusing on hard samples delivers the biggest gains, with improvements of 30–40 points on reasoning benchmarks. How we define hardness: the base model’s pass@k success rate. Easy = solved most of the time Hard = solved rarely, but not never Across GSM8K and BBEH, the hardest 10% consistently outperforms easy, middle, or random subsets. Easy data barely moves the needle, while hard data changes the trajectory of training. Why? GRPO needs reward variance within groups to learn. Easy examples collapse to uniform outcomes quickly. Hard examples sustain a mix of successes and failures, keeping the learning signal alive. And this generalizes: on AIME-2025 (a much harder OOD benchmark), only the hardest subset training improves over the base model, with a 20% relative lift. Takeaway: If you’re doing GRPO post-training under a budget, prioritize the hardest examples. They are where the richest learning signal lives. cc: Benjamin Pikus Burton Ye Paper: https://lnkd.in/dD4aY8kZ Code: https://lnkd.in/d28ezREx We're working on many related, equally interesting problems at Eternis! Hiring for multiple roles: https://lnkd.in/djbEENuW

44

3 Comments

https://lnkd.in/efBuPQMn Is this the Horse to Back I recently came across a well-reasoned piece by Martin Milani critiquing Yann LeCun JEPA (Joint Embedding Predictive Architecture). Milani argues that while JEPA is a master of efficient perception, it lacks the explicit state, causal operators, and counterfactual machinery required to be a true "World Model." In his view, JEPA advances efficiency, not cognition. As a pragmatist looking at the future of telecommunications, I think Milani is right about the classification—but perhaps misses the massive opportunity that this "limitation" creates. We’ve spent years watching LLMs predict the next word in a sequence with remarkable fluency, yet they remain prone to hallucinations because they lack a grasp of physical reality. LeCun’s alternative, JEPA, ignores "pixel-perfect" detail to learn abstract representations of how the world works. If we view the future "Intelligence Stack" as a split system, the missing building block isn't a single, all-knowing model—it’s the network itself. (1) The Great Translator: - In my work consulting for telcos on the impact of AI, I’m increasingly convinced that the 6G network will become the "nervous system" where these models finally find their purpose. If a Mission-Critical World Model (System 2) is the Executive, then JEPA is the Sensory Cortex. The "symbol grounding problem" has long plagued logically perfect systems; they struggle to digest the messy, high-entropy noise of the physical world in real-time. This is where I believe that 6G and JEPA become codependent: JEPA as the Universal Semantic Interface: Its role isn't to provide fiduciary logic, but to act as a high-speed filter. It strips away surface-level noise and converts physical reality into "logic-ready" Latent Embeddings. Efficiency as an Enabler: - By handling the 99% of "intuitive physics" (the reflexes), JEPA allows the formal, heavy causal reasoning systems to focus only on the most relevant variables. (2) The Bridge to 6G: - The future isn't about JEPA evolving into a complete, autonomous reasoning agent, rather JEPA becoming the interface that allows logically grounded systems to "feel" and "react" to the physical world at scale. In a 6G-native environment, the network doesn't just move data; it processes perception / reacts to logic. JEPA provides the perceptual intuition, while the network provides the connectivity for the "fiduciary logic" of governance and trust to intervene when it matters most. I would like to think that we don't need JEPA to be the brain; we need it to be the bridge.

12

2 Comments

We cut the hallucinations of an LLM in half using a new technique: RLFR (Reinforcement Learning from Feature Rewards). RLFR reduces hallucinations in Gemma 12B by 58% (when run with our probing harness), for ~90x cheaper per intervention than the LLM-as-judge alternative. The technique uses lightweight probes on a model's internal representations as reward signals for reinforcement learning. This is all while avoiding: - off-target effects (no degraded performance on standard benchmarks) - degrading/reward-hacking the probes (the probes still work as monitors at test time!) This is just our first public demonstration of our work building towards "intentional design": the ability to intelligently guide gradient descent, enabling a new paradigm of far more precise, robust, and effective training. Kudos to Aaditya Prasad, Connor Watts, Jack Merullo, Dhruvil Gala, Owen Lewis, Tom McGrath, and Ekdeep Singh for this groundbreaking new work!

234

9 Comments

From Precision to Exploration: Many developers still try to make AI behave like a compiler, when it’s really more like a creative partner. Developers are trained to think in right and wrong. But AI tools don’t live in that world. When you work with LLMs, the question isn’t “Is this correct?”. It’s “Is this good enough to refine?” It’s a world of probabilities, not absolutes. The best developers I’ve seen don’t try to force AI into binary logic. They iterate, guide, and sculpt, and they do it fast. They understand that quality emerges through iteration, not prediction. You don’t build perfection with AI. You discover it, piece by piece. 👉 I challenge you to think: Are you still coding for correctness, or designing for evolution? #AI #QA #DigitalQuality #SoftwareDevelopment #Testing

5

1 Comment

Yann LeCun is leaving Meta, and many people have linked his departure to his criticism of LLMs. But what exactly is he criticizing? I’ve seen a lot of speculation, so let’s take a closer look at what he actually says. First, LeCun is not claiming that LLMs are useless. On the contrary, he has repeatedly stated that they are valuable and powerful systems. His argument is more specific: auto-regressive LLMs that are predicting the next word are not the path to human-level intelligence. According to LeCun, truly intelligent systems need four key capabilities: 1. Understanding (a representation) of the physical world     2. Persistent memory   3. Ability for Reasoning   4. Ability for Planning  LLMs don't excel in any of these. They have some of these properties, but only in very limited forms. Why is that the case? LLMs are trained solely on text. They are basically statistical parrots that predict the next word of a sentence. That predicted word is then added to the sentence as input for the next prediction. This process continues (auto-regressively) until the full response is generated. For progress toward superintelligence, this method has clear limitations: - LLMs are basically “mumbling” a response on the fly, word by word, without forming an internal plan of what to say. Humans don't act like this, they think before they speak.   - Text is low bandwidth compared to the sensory data humans absorb from the physical world. We learn most of what we know of the world by interacting with our environment, not by reading about it. A four-year-old has processed more visual information than what is contained in all human text ever written.   - LLMs can pass difficult exams, they still struggle with tasks that are trivial for humans, like clearing a table, loading a dishwasher, or driving a car.  LeCun’s argument is straightforward: predicting the next word will not get us to systems with deep physical understanding. To reach human-level or superintelligent capabilities, we need models that embody the four properties above, not just better text prediction. I agree with him on that. But this is not a binary situation. I think that LLMs are really useful AND I think they are not the way to human level intelligence. I highly recommend this conversation between Yann LeCun and Lex Fridman from last year. This will make things a lot more clear: https://lnkd.in/e9bgNRK7 Image source: https://lnkd.in/egUwNpSs

27

8 Comments

"Can LLMs be Computers?" This paper absolutely blew me away: https://lnkd.in/dJTS76Ts The entire agentic ecosystem today, from Claude Code to OpenClaw, is based on the principle of surrounding LLMs - which are simply token distribution generators - with external tools (Python, git, curl, APIs, MCPs), and a state machine. An LLM cannot reliably "execute" a mathematical expression - it can talk/reason about it; it can (if seen in the training data) "guess" the result; it can produce Python code to run (externally) in a Python sandbox, and use the result in its answer. What the folks in this paper did, was turn the LLM into a virtual machine (VM), and effectively use the transformer weights as a fast WebAssembly interpreter. The result: you give an LLM a program, it ACTUALLY executes the program by emitting a stack trace (internally), and it outputs the result! They took a vanilla PyTorch Transformer architecture (couple of dozen lines of code), and embedded WebAssembly in model weights by training the transformer to predict state transitions of a WASM virtual machine, so the attention and MLP layers collectively implement the interpreter’s logic. Programs are fed as tokens, and the network simulates execution step-by-step. If it holds true, this could be an absolute game changer. Think of one agent turning written/spoken instructions into code, and feeding it to another agent that is acting as a VM/interpreter - BUT now you can do it all in one go (single agent)! Vilhelm von Ehrenheim I thought of you when I read this!

136

18 Comments

We’ve spent the last several months working closely with customers building real agents - research agents, coding copilots, multi-step workflows with tools. The pattern was clear: traditional RL setups don’t match how production agents actually behave. Today we’re introducing Fireworks RFT, a reinforcement fine-tuning system built around real agent trajectories: multi-turn, tool calls, retries, everything. Instead of asking teams to redesign their agents for RL, we plug directly into the workflow they already have. The early results have been impressive - agents that outperform closed models, use tools more effectively, and run at dramatically lower cost. If you’re working on agents for research, coding, structured reasoning, or domain workflows, this is a meaningful step forward.

64

2 Comments

🚨 The AI world just shifted on its axis. Yann LeCun—the Turing Award winner who helped build the foundation of modern AI—just launched Advanced Machine Intelligence (AMI) with a bold €3 billion valuation target. This isn't just another AI startup. This is a direct challenge to the entire LLM paradigm. While everyone's obsessing over ChatGPT and language models, LeCun is betting on something completely different: "world models." Here's why this matters: 🧠 World models don't just process text—they simulate entire environments and understand cause-and-effect relationships. 🎯 They're designed to eliminate the hallucination problem that plagues current AI systems. ⚡ This could be the breakthrough that takes AI from "impressive chatbot" to "true machine intelligence." The timing is fascinating. Just as the industry starts questioning whether we've hit a wall with LLMs, one of AI's founding fathers is saying: "We're looking in the wrong direction." With Alex LeBrun (Nabla's co-founder) as CEO and LeCun as Executive Chairman, AMI is reportedly seeking €500M in funding to make this vision reality. This could either be the next paradigm shift in AI—or the most expensive pivot in tech history. What's your take: Are we ready to move beyond language models, or is LeCun betting against the wrong horse? #AI #MachineLearning #Innovation #TechNews #ArtificialIntelligence

6

The upcoming upgrade to the teckel app facilitates natural, flowing conversation with multiple LLMs at the same time. Here's an example transcript: https://lnkd.in/e4kgxJiH I start by asking GEMINI a few direct questions whilst PERPLEXITY, DEEPSEEK, and CLAUDE listen in the background without commenting (we call this "CrossListen"). Then I ask CHATGPT, PERPLEXITY, DEEPSEEK, and CLAUDE to respond to what GEMINI has just said. Finally, I ask them to reach consensus amongst themselves (we call this "CrossTalk"). (This version of the app is in beta-testing and will be published soon.)

5

1 Comment

Thrilled to share the news about the newly founded lab at Imperial College London funded by Thomson Reuters that will fund new PhD students and research on foundational ML, safety and societal impact.

11