The key distinction: More test-time compute is not automatically useful. It becomes useful when the model has learned an internal landscape where extra iterations move the latent state toward solution-aligned attractors rather than spurious ones. That is why the convergence
RESEARCH
-
OpenAI and Anthropic’s contrasting AI launches in 2026 cinema
By
–
OpenAI: carefully rolls out GPT-5.5-Cyber through Trusted Access for verified defenders Anthropic: “Claude Mythos is too powerful for public release” Also Anthropic: accidentally shows Mythos in the UI and immediately runs out of capacity 2026 AI launches are absolut cinema.
-
Open-source AI wins through community and knowledge sharing
By
–
I said Buy a GPU because I wanted as many indie researchers & devs working on RTX 3090s for the community (e.g. @pupposandro
, @no_stp_on_snek
) I started writing greentext posts on LLMs because I wanted to get people curious How Opensource AI wins? Community & knowledge sharing -
Open source AI must stay independent and community-governed
By
–
Opensource AI should remain usable, understandable, reproducible, locally deployable, economically viable, and community-governed even if today's dominant labs, foreign labs, hardware vendors, cloud platforms, or open-weight model providers change direction or disappear.
-
Neurosymbolic Necessary but Not Sufficient, Says Marcus
By
–
neurosymbolic is necessary but not sufficient, see eg my 2020 arxiv The Next Decade in AI and the 2018 arxiv with innateness in the title
-
Marcus agrees on need for intrinsic spatial and temporal logic
By
–
oh i certainly agree on the need for intrinsic spatial and temporal logic, see eg my 2018 essay with innateness in the title, rebooting AI, etc
-
Data Value Density framework boosts AI learning from less data
By
–
What if AI could learn more from less data? Researchers from Shanghai Jiao Tong University and Shanghai AI Lab introduce 'Data Value Density (DVD) enhancement' — a unified framework to make every training token count. Instead of just piling on more internet data, DVD methods
-
Cloud codex runner using Cloudflare Firecracker and Ghostty
By
–
Still limited by compute, so I built a thing that runs codex in the cloud, powered by @Cloudflare firecracker boxes (and since that's not beefy enough for larger projects, tests are run via crabbox) Uses Ghostty ofc, via WebAssembly.
Codex replicated itself, basically. -
ConvexTok: Linear Programming for Optimal LLM Tokenization
By
–
"Tokenisation via Convex Relaxations" Most LLM tokenizers still use BPE, a greedy merge algorithm that can waste vocab slots on locally good but globally suboptimal tokens. This paper turns tokenizer training into a linear program, then rounds the solution into ConvexTok. This
-
Gary Marcus: Deep Learning hitting a wall, possible Substack essay
By
–
almost all is clearly explained in my article Deep Learning is hitting a wall, in Nautilus: https://
nautil.us/deep-learning-
is-hitting-a-wall-238440
… but maybe i will go line by line in a substack essay?