
What happened
Pohang University researchers Hyunjong Ok and Jaeho Lee report that Claude Opus 5.5, GPT-6.1 Sol and GPT-6 Astra, asked to write neural-network weights directly without gradient training, all converged on sparse lookup tables. The 0.94M-parameter designs scored 54.4% on ARC versus 29.8% for 3.33M-parameter transformers from the same LLMs, with up to 17x faster GPU inference.
Why it matters
The finding suggests that models built by today's frontier LLMs may be far smaller and faster than transformer equivalents while scoring higher on ARC, which could change how teams think about model size and inference cost.
WHO IT HITSMachine-learning researchers and engineers who design or deploy compact models may need to reassess whether gradient-trained transformers are the default architecture; the result points to sparse lookup tables as a viable alternative.
Summaries like this, every morning.
The result comes from a deliberately unusual setup: instead of training a network with gradient descent, the researchers had three frontier models write the weights directly. All three independently landed on sparse lookup tables rather than transformers, which suggests the convergence is not tied to one model family.
The performance gap is striking. Despite having roughly a third of the parameters of the transformer baselines produced by the same LLMs, the lookup-table models scored 54.4% on ARC versus 29.8%. The speed difference is also large, with up to 17x faster GPU inference.
A practical consequence is traceability: because every prediction maps back to a written fact row, local edits change only about one answer in 3,530. That property may matter for teams that need to inspect or adjust model behavior without retraining.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Baidu said on September 21, 2026 that Kooko, the overseas edition of its Kooko AI office assistant, passed 10…

FastMetal started offering Qwen4 Max on Nov 17, 2026, a day after Alibaba published it; the model has 2.4 tril…

Snapchat introduced Easy Lens CLI, which lets AI assistants like Claude Code, Cursor and Codex build Lenses fr…

Salesforce wired Amazon SageMaker AI's cross-Region inference routing into its Internal Gateway, so one manage…

OpenAI said Apollo Research, METR, Transluce and Goodfire will work on-site at its offices, with access to unr…

Walt Disney Imagineering R&D revealed three Audio-Animatronic prototypes—Vilo Vele, Kado Bre'a, and Jibu Bre'a…
