there’s a whole article on that phenomenon, in fact https://
garymarcus.substack.com/p/the-mirage-o
f-visual-understanding?r=8tdk6
…
LLMS
-
The Mirage of Visual Understanding in AI
By
–
-
Autonomous AI Agents: Hacking and Self-Replication
By
–
A few days ago, Palisade Research confirmed that AI agents can now autonomously hack remote computers and self-replicate, with success rates jumping from 6% to 81% in just one year. In tests, a Qwen 3.6 agent navigated across four countries, installing its own weights and launching operations.
-
Debating the reach of AI capabilities beyond math and coding
By
–
oh i agree on math and coding; but he is saying it will go far beyond to open-ended domains, and that’s where the disagreement lies
-
Analyzing Model Accuracy Thresholds and Task Performance
By
–
they have tasks out to 16 hours (i believe_, so there is plenty of headroom at 90% accuracy. which is to say lots of tasks in the current edition of the task where the model is not at 90%. if you insist on 95% accuracy even more. the 50% is just an arbitrary criterion; there
-
OpenAI Codex Unveils New Ultrafast Mode for Latency-Sensitive Work
By
–
OPENAI : A mention of a new Ultrafast mode appeared for some time on the Codex GitHub repository. > "The fastest available responses for latency-sensitive work." Seems like it was unintended push
-
Local open-weight AI outpaces Moore’s Law twofold
By
–
Local open-weight AI running on a laptop has been improving at more than twice the rate of Moore's Law! Between May 2024 and May 2026, the most expensive MacBook Pro available still capped at 128 GB of unified memory. The hardware ceiling barely budged. Yet the most advanced open-weight model…
-
Deconstructing Geoffrey Hinton’s Views on AI
By
–
the details of what he said and why it was wrong are here: https://
open.substack.com/pub/garymarcus
/p/deconstructing-geoffrey-hintons-weakest?r=8tdk6&utm_medium=ios
… -
Enterprise roadmap vs Labs’ rapid AGI scaling vision
By
–
Enterprises are going to actually want a coherent roadmap for the development of tools like Codex and Cowork, so they can plan and train and scale their use. This conflicts with the Labs’ vision where these tools rapidly scale exponentially in ability as models approach AGI.
-
Essential Technical Skills for AI Engineers
By
–
As an AI Engineer. Please learn: – Harness engineering, not just prompt engineering
– Prompt caching vs. semantic caching tradeoffs
– KV cache management at scale
– Speculative decoding vs quantization
– Structured output failures & fallback chains
– Evals (LLM-as-judge + human -
Technical Requirements for Evaluating AI Model Architecture
By
–
Thanks, but there are no open weights yet, right? Asking because it would be impossible to cover architecture details without open weights and/or a detailed technical report