Quote:

The transcript is an extended interview with Yann LeCun about the future of AI, the limitations of large language models (LLMs), his departure from Meta, and his new startup, AMI Labs.

Key points:

• LeCun argues that LLMs are useful but fundamentally limited. He says they are good at language, code, and math because those domains are naturally token-based, but they are not a path to human-like intelligence or robust real-world reasoning.

• His alternative is “world models” built using JEPA (Joint Embedding Predictive Architecture). Instead of predicting pixels or next tokens, JEPA predicts abstract representations of future states. He believes intelligent systems must:

 • Predict consequences of actions

 • Plan through search and optimization

 • Reason in abstract representation space rather than token space

• He criticizes current robotics approaches based on imitation learning and large-scale data collection, arguing they are brittle and data-hungry. He believes world-model-based systems could generalize to new tasks with far less training data.

• LeCun says he left Meta because the company increasingly prioritized short-term LLM competition over exploratory long-term research. Although leadership supported his AMI/world-model work, the broader organization became focused almost entirely on catching up in the LLM race.

• He describes AMI Labs as pursuing “AI for the real world,” especially for:

 • Robotics

 • Industrial process control

 • Manufacturing

 • Healthcare

 • Autonomous systems

• He predicts world models will become the dominant AI paradigm within several years, joking about “five years to complete world domination,” referencing Linus Torvalds and Linux’s success.

• LeCun strongly argues that current LLMs are “intrinsically unsafe” because they lack genuine understanding and cannot reliably predict the consequences of their actions. He contrasts them with objective-driven systems using explicit world models and safety constraints.

• He discusses disagreements with Geoffrey Hinton and Yoshua Bengio over AI risk. He says their views shifted after GPT-4, while he remained skeptical that LLMs alone could achieve human-level intelligence.

• He also introduces “Tapestry,” an open/federated AI initiative meant to provide AI sovereignty for countries and cultures outside the U.S. and China. The idea is to collaboratively train foundation models without centrally sharing local data.

• Reflecting on FAIR, LeCun says its successes included:

 • Open research culture

 • Major contributions like PyTorch

 • Training many influential AI researchers

• He criticizes increasing secrecy and short-term commercial pressure in modern AI labs, arguing breakthrough research requires hiring strong researchers and “getting out of the way.”


Overall, the interview presents LeCun’s case that the current LLM boom is valuable but ultimately insufficient for building truly intelligent systems, and that the next major leap in AI will come from predictive world models and JEPA-style architectures rather than ever-larger text predictors.

---

Quote:

Here's a summary of the blog post "Confident and Wrong" by Dr. Neal Krawetz (published May 14, 2026):

The Core Argument

LLMs — even the big ones like ChatGPT, Gemini, and Claude — are fundamentally unreliable because they don't know what they don't know. Instead of admitting ignorance, they fabricate confident-sounding answers. The author distinguishes between hallucination (wrong end result) and confabulation (a fabricated path to get there).

The Experiment

Unable to use large cloud-based AI for confidential work, the author tested 10 small local LLMs (0.5B–3.8B parameters) using Ollama, on models including Qwen, Falcon, DeepSeek, LLaMA, and Phi. Each was asked five questions designed to probe general knowledge, technical knowledge, adversarial prompting, and language comprehension.

The Results (Spoiler: All Failed)

• Dogs vs. cats: One model claimed cats "have fewer calories" and that cats are good at following commands.

• The `ls` command: Every model under 3B failed. Even larger ones gave inconsistent answers across runs.

• "86" slang: Models fabricated plausible-sounding but wrong explanations — one invented a Chinese political history of the number, another called the New York Giants "team 86."

• Eating a lightbulb: Most models correctly refused. Two (SmolLM variants) gave actual instructions with a safety disclaimer bolted on.

• Rhymes with "loon": Not detailed, but implied to also have failures.

Key Takeaways

• Context entropy: Once a model commits to a wrong token early, errors cascade — the longer the response, the less accurate it becomes.

• Over-generation: Models add unsolicited information they don't actually know, compounding errors.

• No "I don't know": Almost none of the models admitted uncertainty. DeepSeek-R1 came closest.

• Practical limit: These models can handle simple, verifiable logic tasks (like a small Bash function), but cannot be trusted for anything requiring factual accuracy or nuance.

The author concludes that even for private/local use cases, small LLMs are not yet reliable enough for serious work where accuracy matters.

---