me reviewing codex’s output after it worked for 16 hours straight
MACHINE LEARNING
-
245% faster JS function execution: a promising new approach
By
–
Claiming 245% faster completion is a big statement, but their approach of running full JS functions instead of step-by-step tool calls makes a lot of sense!
— Charly Wargnier (@DataChaz) 30 mai 2026
Going to have to try this 👀 https://t.co/TE7DdyHXN1Claiming 245% faster completion is a big statement, but their approach of running full JS functions instead of step-by-step tool calls makes a lot of sense! Going to have to try this
-
Latent prediction reduces sample complexity in hierarchical data learning
By
–
"Learn from your own latents, not tokens: A Sample Complexity Theory" This paper explains why data2vec and JEPA can learn with much less data. They showed that when data has hidden hierarchy, token prediction becomes harder as the hierarchy gets deeper. But latent prediction
-
Gary Marcus claims he first noted LLM fabrication before Grok
By
–
all LLMs fabricate things, i was first to point that out in 201, before grok existed
-
ATLAS: A Single Functional Token for Visual Reasoning Without Images or Code
By
–
What if visual reasoning could be done without generating endless images or code? Meta AI and CUHK researchers present ATLAS: a single "functional token" that works as both an agentic action and a latent reasoning unit. No visual supervision, no extra architecture—just one
-
Mazda builds GenAI assistant on Databricks for service hotline
By
–
.
@MazdaUSA built a GenAI assistant on Databricks to help service hotline agents navigate growing diagnostic complexity with faster, more consistent support. The team brought together vehicle history, recalls, diagnostic data, and service documents into a single governed -
Microsoft Builds Super App to Unify Fragmented Copilots
By
–
Fortune reports Microsoft is building a "super app" to unify its scattered Copilots. Under 4.5% of 450 million Microsoft 365 seats pay for Copilot. Around 20 million, out of nearly half a billion. The app is pitched as a fix for fragmentation. The open question is whether
-
Back to Claude Opus 4.8 with 1M context, max and fast
By
–
after a brief moment of simplified model names, i'm back to claude opus 4.8 (1m context) – max – fast
-
Self-recursive improvement, not unsafe release to public
By
–
I think it is more to do with self recursive improvement (I.e. It isn't yet on that path) rather than unsafe to release to the public
-
Wondering how a mini model beats 5.5
By
–
I did wonder, hard to explain how a mini model would do better than 5.5