Max
6K posts
Max
@MaxScore
building machines that understand what they see | @manakoai | @webuildscore ⎸ sire ⎸ opinions are my own
Just got a masterclass from @MaxScore on computer vision and training robots: - Most robot-training techniques bottom out in collecting visual data, hence the flood of capital into training facilities and camera-headset human-recording rigs - Manako started from a reasoning model and taught the model to see, rather than starting from a vision model and bolting on reasoning - Manako's VLA is a distilled reasoning model + a traditional VLM + a pure detection model, running ~10 bespoke primitives. - The world is shifting toward World Models: "being intelligent is not about what you know, it's about what you do when you don't know" - This is broadly supported by firms like Physical Intelligence; Predictive (LeCun / JEPA-style joint embedding predictive architectures): you don't need to see at all, you predict embeddings or latent states. Max is fascinated by this Vision is not solved, we are waiting for the Transformer moment we saw with LLMs. Max would like to contribute to progressing research here via Manako and their partners
pushed a big autofrontier 'recursive self-improvement' update tonight. the loop now improves itself at three levels: level 1, it already did: train better models overnight, a strict gate deciding which gains are real. level 2, new: the gate also tests the strategy used to find those gains. better strategies kept, worse ones dropped. level 3, new: the loop can improve its own code, same gate judging. so one night of GPU gets you a better model AND a better way of finding the next one. every night compounds. repo in comments.
putting autofrontier to work for real: training satori-8B base. the frontier: referring-expression grounding (Acc@0.5) on RefCOCO. genesis: 0.850 bar to break: 0.927 (Qwen2.5-VL-72B) an 8B chasing a 72B's record, agent hunting angles overnight on one GPU. 3 duels in: 1 crowned → king now 0.920 2 rejected → one posted 0.95, above the bar, and still refused: the edge didn't replicate repo in comments.
Latency update, first results. Four production-ready models live on SN44 right now: CRIME: 5MB, 70-80ms, ~75% accuracy FIRE: 9.8MB, 60-70ms, ~90% accuracy CARWASH: 9.7MB, 60-70ms, ~75% accuracy ROADSIGN: 9.8MB, 70-80ms, ~85% accuracy Read those numbers again 👇 Under 10MB. Under 80ms. Production-grade accuracy. Small enough to run on a camera, a gateway, an embedded board... That's the frontier we're optimizing: accuracy per byte, per millisecond. The latency loop enforces it, and miners are converging faster than we expected. More elements coming.
Just got a masterclass from @MaxScore on computer vision and training robots: - Most robot-training techniques bottom out in collecting visual data, hence the flood of capital into training facilities and camera-headset human-recording rigs - Manako started from a reasoning model and taught the model to see, rather than starting from a vision model and bolting on reasoning - Manako's VLA is a distilled reasoning model + a traditional VLM + a pure detection model, running ~10 bespoke primitives. - The world is shifting toward World Models: "being intelligent is not about what you know, it's about what you do when you don't know" - This is broadly supported by firms like Physical Intelligence; Predictive (LeCun / JEPA-style joint embedding predictive architectures): you don't need to see at all, you predict embeddings or latent states. Max is fascinated by this Vision is not solved, we are waiting for the Transformer moment we saw with LLMs. Max would like to contribute to progressing research here via Manako and their partners
GPT-5.6 Sol just one-shotted this Chess app that can record your OTB games in real time and it’s insanely good 🤯🤯🤯
“The model alone is no longer the product" @AravSrinivas says the real product is now the harness around it: orchestration, tools, enterprise context, and cost performance. The post-frontier AI race is about systems, not just the smartest model
