Yann LeCun's $1B Bet Against LLMs

Credibility score: 81/100 — Highly Credible. This video is highly credible with well-supported claims.

Claims analyzed

Yann LeCun raised $1B for non-generative AI called Jeepa/Jeppa — Sketchy (40/100)

Names butchered + $1B claim with zero receipts. Smells like hype 🚩💀

JEPA is alternative architecture using embeddings — Solid (85/100)

👌

JEPA uses encoders and predictor on embeddings — Verified (95/100)

✅

LLMs good at language manipulation, nothing else — Opinion (50/100)

LeCun's classic LLM hot take — spicy but debatable 🔥

LeCun pioneered CNNs in 1980s while others built expert systems — Solid (85/100)

👌

AlexNet similar to LeCun's 1990s CNNs, 25 years later — Solid (90/100)

✅

RL renaissance mid-2010s via DeepMind Atari and Go — Verified (95/100)

👌

OpenAI founded 2015, focused on RL with Gym/Universe for games — Verified (95/100)

👌✅

LeCun's famous cake slide meme from ~2015 — Personal Story (50/100)

Classic cake slide drop — RL as cherry is savage 🔥

Radford modified Transformer for next-token prediction self-supervision — Solid (90/100)

👌

GPT trained on 7,000 books dataset, then supervised fine-tuning — Solid (88/100)

✅

GPT set SOTA on 9 benchmarks incl. reading comprehension, beat task-specific models — Verified (92/100)

😤✅

Radford's model is GPT-1 — Solid (85/100)

👌

GPT training matched Yann LeCun's predictions — Dubious (45/100)

LeCun predicted *exactly* this pipeline? That's a stretch 🚩

LeCun tried video prediction years before GPT-1 — Solid (75/100)

✅

Autoregressive video prediction gets blurry fast — Solid (85/100)

✅

Hudson River Trading sponsor - order book demo — Sponsored (50/100)

Mid-video sponsor plug 💼

Hudson River Trading sponsor read and hiring pitch — Sponsored (50/100)

Classic mid-video sponsor plug — smooth transition tho 👌

GPT-2 has 50,257 tokens in vocab — Verified (100/100)

✅

Full HD frame has ~10^15M possible values — Solid (85/100)

👌 Math checks, exaggeration for effect

Video models predict average frame, causing blur — Solid (85/100)

👌 Classic mode collapse demo

Next token prediction works shockingly well as intelligence proxy — Opinion (50/100)

'Shockingly well' — yeah, till it hallucinates your grandma's recipe 💀

Need other methods beyond next-token for intelligent systems — Opinion (50/100)

Pumping the brakes on LLM hype — fair play 😏

Best image representation systems since 2017-18 are non-generative — Solid (85/100)

✅

He's worked on joint embeddings/Siamese nets since 1990s — Verified (95/100)

👌

Siamese nets created in early 1990s for fraudulent signature detection — Verified (100/100)

✓💯

Siamese network learns useful representations without generating images — Solid (85/100)

👌

Joint embeddings sidestep blurry video issues of generative models — Solid (80/100)

✅

Joint embedding risks trivial constant output solution — Verified (95/100)

💯

LeCun's Siamese networks used contrastive learning to avoid representation collapse — Verified (95/100)

✅

See the full analysis with sources and timestamps →