Ask not where AI is, but where it will asymptote.
RESEARCH
-
Unrealistic AI Hype and Follower Disappointment Cycles
By
–
Se sobre-emocionan ante la expectativa de algo que no tiene sentido, y luego se sobre-decepcionan cuando esas expectativas falsas no se cumplen. Y con ellos sus seguidores.
-

AI Influencers Misunderstand METR Benchmark Results for Sonnet
By
–
The number of AI influencers who are surprised that Sonnet 4.5 didn't achieve a better position on the METR benchmark, when they were saying it "could work autonomously for 30 hours," worries me. They're not understanding anything about what these benchmarks measure. They're
-

Gemini 2.5 Deep Think Achieves State-of-the-Art on FrontierMath
By
–
Gemini 2.5 Deep Think is SoTA on FrontierMath!
-
Build Autograd Engine, Mini-GPT, and LoRA Fine-tuning From Scratch
By
–
– build an autograd engine from scratch
– write a mini-GPT from scratch
– implement LoRA and fine-tune a model on real data
– hate CUDA at least once
– cry
– keep going the roadmap – 5 phases
– if you already know something? skip
– if you're lost? rewatch
– if you’re stuck? use -

Inference Efficiency Drives AI Production Profitability and ROI
By
–
As AI moves from prototypes to production, inference efficiency becomes the ultimate driver of profitability. Open benchmarks like InferenceMax provide a shared standard to measure performance, TCO, and ROI, validating NVIDIA’s full-stack approach of co-designed hardware and
-

DeepThink IMO Gold Model Achieves State-of-the-Art FrontierMath Results
By
–
Very cool to see this new state-of-the-art result on FrontierMath achieved by the #DeepThink IMO Gold model that we built 3 months (a long time) ago 🙂 The fact that is evaluated externally using a publicly available API is a strong testament that what we built generalizes beyond
-
Alphabet Showcases Four Major AI-Driven Innovations Including Genie 3
By
–
Genie 3 is one of 4(!) Alphabet innovations featured, including FireSat (satellites that use AI to better detect wildfires), CiviForm (open-source software that consolidates public program applications into one platform) & @Waymo (autonomous vehicles). Congrats to all the teams!
-
AI and analog computing revolutionize transformers
By
–
Sources:
1/
https://research.ibm.com/blog/how-can-analog-in-memory-computing-power-transformer-models?utm_source=chatgpt.com
… 2/
https://nature.com/articles/s43588-025-00854-1?utm_source=chatgpt.com
… -

Multiple RL Trajectories Improve Agent Strategy Diversity
By
–
We have always been using 1 RL trajectory for training. Why not use more? In this research: Polychromic Objectives for RL, they train their model on sets of trajectories and reward both success & diversity, so the agent keeps MULTIPLE strategies alive all at once.