ScaledML is back! After a 6 year hiatus. Consistently 2-3 years ahead of where ML will be. Examples of foresight at SML:
– OpenAI in 2016,2017,2019 announced GPT-2 & RL efforts
– Turing award for Deep Learning announced by Turing award winner on morning of award
– Groq chip
MACHINE LEARNING
-
ScaledML Returns After 6 Years with ML Industry Foresight
By
–
-
Modern Deep Learning Detectors: Why HOG Is Outdated
By
–
Still using HOG for object or pedestrian detection? You probably shouldn’t.
— Satya Mallick (@LearnOpenCV) 7 janvier 2026
HOG was great years ago, but today it’s slow, fragile to viewpoint changes, and struggles with small or occluded objects.
Modern deep learning detectors in OpenCV — like U-Net (faces), YOLO, and… pic.twitter.com/YM3oDyF0XXStill using HOG for object or pedestrian detection? You probably shouldn’t. HOG was great years ago, but today it’s slow, fragile to viewpoint changes, and struggles with small or occluded objects. Modern deep learning detectors in OpenCV — like U-Net (faces), YOLO, and
-
AI-driven detection of cyberharassment using graph studies
By
–
Oui, c’est un move intéressant. Je dis « réducteur » parce que les chercheuses de l’I3S de Nice m’avaient expliqué que la détection du cyberharcèlement repose en grande partie sur des études de graphes et donc sur des données massives. Et oui, elles ne pouvaient pas accéder aux
-
Keras API: Accessing JAX’s Full Power Without Reinventing
By
–
The point of the Keras API is to let you tap into everything JAX has to offer but without reinventing the wheel
-
Technical evaluation of AI-based harassment detection systems
By
–
J’update mon message : apparement c’est basé sur la lecture des messages. Et non pas sur de l’harcèlement publique donc. Donc on détecte uniquement le harcèlement des messages. Ok c’est un choix. Réducteur, mais un choix. D’un point de vue technique c’est ok, mais ça veut
-
Limitations of AI-based automated detection in current platform environments
By
–
Personne ne me demande mon avis, mais je le donne gratuitement. C’est une bonne idée, mais malheureusement avec 9 ans de retard. La tech des plateformes aujourd’hui ne permet plus d’automatiser ce type de détection avec de l’IA. Seuls des chercheurs ont accès aux données, sous
-
Structured Plans and Decision Theory Maximize Model Efficiency at Scale
By
–
6/6 This is the core takeaway: With structured plans and decision-theoretic optimization techniques, you can get a lot more with the same models, tools and compute. And as you scale, the gap widens.
-
ReAct Agents and High-Variance Execution in Long-Horizon Tasks
By
–
4/6 ReAct-style agents encode the entire plan in text, and only get evaluated at the very end. In long-horizon tasks, that turns execution into a high-variance gamble.
-
Maestro: Optimizing Compute Through Agentic Framework Structure
By
–
2/6 SWE-bench is not the story. Maestro is not a dedicated SWE agent, it’s an agentic framework that optimizes based on structure and learned priors. ex: knowing that 4 runs of GPT-5 mini are cheaper and better than one GPT-5 run, leads to better use of the same compute –
-
Test-Time Compute: Optimizing Where and When to Spend Resources
By
–
3/6 Once you treat execution as a test-time compute problem, the question changes. It’s no longer “which model should I use?” It’s “where should I spend compute, and when should I stop?”. That’s why multiple cheaper runs can win:
you’re buying optional paths, not just more