Looks like DeepSeek is handing hardware optimization control directly to developers in the latest DeepGEMM update. For the fp8_mqa_logits function, the weights tensor dtype now explicitly dictates the accumulation precision.
AI
-
Using Codex to drive OpenClaw: a correction
By
–
it’s the other way around, you can use codex to drive OpenClaw.
-
Microsoft déploie sept modèles IA spécialisés en une fois
By
–
Wild to see Microsoft roll out seven specialized models in one shot-reasoning, coding, image gen, transcription, and voice all covered in a single product-ready stack. This is what it looks like when AI stops being a demo and starts becoming infrastructure. If you care about
-
Bet with Miles Brundage: AI not smarter than smartest humans
By
–
see also my bet with Miles Brundage. lots of things smart people can do that AI still can’t. the claim was smarter than smartest human, and they have not met that.
-
Microsoft Builds Own AI: 7 New Models Trained From Scratch
By
–
🚨Microsoft just stopped renting intelligence and started building their own!
— Futurepedia – Learn to Leverage AI (@futurepedia_io) 3 juin 2026
7 new MAI models. Reasoning. Coding. Image. Voice. Transcription. All trained from scratch. Zero distillation. No third-party model outputs. Just clean data and their own infrastructure.
Their… pic.twitter.com/kilTDSO8lNMicrosoft just stopped renting intelligence and started building their own! 7 new MAI models. Reasoning. Coding. Image. Voice. Transcription. All trained from scratch. Zero distillation. No third-party model outputs. Just clean data and their own infrastructure. Their
-
AI alignment EA adjacent but rationality failed morality
By
–
Good AI alignment is necessarily EA adjacent. But EA failed to derive a global morality, not for lack of good intentions or intelligence, but because human rationality is a brittle tool, even if you are a good rationalist, and most people prefer to either muddle through or to
-
14 RTX 3090s running 42 parallel agents at full 256k context
By
–
14x RTX 3090s + Qwen 3.6 27B Running 42 agents IN PARALLEL at full 256k context – exl3 6bpw
– fp8 KV Cache
– Aphrodite Inference Engine w/ tp=2, pp=7 The world of agents will run locally btw -
AI reads 40k small accounts to build news page
By
–
It is true. Which is why I bug lists if about 40,000 small accounts. “AI Community” lists are ask people with fewer than 20,000 followers. AI Newsmakers and AI Influencers are big. https://
x.com/scobleizer/lis
ts
… Then I made an AI to read them all and build a news page of the best: -
Power, soundness, tractability: the AI trilemma
By
–
Power, soundness and tractability: pick any two.
Sacrifice power: statistics
Sacrifice tractability: symbolic AI
Sacrifice soundness: neural networks -
Build and launch apps to your team using Codex
By
–
Build and launch apps to your team, using Codex: https://t.co/eyR0uDDbsq
— Greg Brockman (@gdb) 3 juin 2026Build and launch apps to your team, using Codex: