ai-powered-markdown-translatorArticle translated from fr to en with gpt-6.1-sol.
On the Hugging Face community blog, Abhishek Sensharma publishes a world model (world model) that generates GTA V frame by frame from player keystrokes, with its weights under Apache-2.0. For coding agents, Devin makes every MCP tool configurable, Qwen Code v0.27.4 gives its managed agent access to local folders, and a Codex CLI preview passes cost and usage limits to status lines provided by an MCP server. A Naxi translation model and a hidden verifier presented at the Lexsi Labs hackathon round out the news.
A 630 million parameter world model generates GTA V frame by frame from player keystrokes
November 15 — Under the name lucidml, Abhishek Sensharma publishes a GTA V world model, a transformer he says was trained from scratch: starting from an initial image and WASD keystrokes, it generates the game frame by frame, without a game engine. Weights and inference code are under Apache-2.0, with a playable demo; training code is not provided. The model keeps the two most recent frames in full and retains only a summary it writes itself for older frames.
| Model characteristic | Value according to the author |
|---|---|
| Parameters | 630 205 008 |
| Summary of an older frame | 96 tokens, versus 1 008 |
| Specialization data | about 800 hours of gameplay |
An individual project, unaffiliated with Rockstar Games. The author notes that the checkpoint was not trained to convergence because of budget constraints, and the model does not yet maintain context for long.
🔗 lucidml’s post (Hugging Face)
Briefs
Coding agents and CLI
- Codex CLI 0.171.0-alpha.6 — The notes for this preview are empty; according to its 73 commits since alpha.5, status lines provided by an MCP server now receive token consumption, context usage, estimated cost and Codex usage limits. The latest stable release remains 0.169.0 from November 13. 🔗 source
- Qwen Code v0.27.4 — Two days after v0.27.3, this stable release adds outbound-only Feishu and WeCom webhooks, gives managed agent (Managed Agent) sessions access to local folders granted through the browser, and has Auto mode check that risky actions have been explicitly authorized. No known breaking changes across 91 entries. 🔗 source
- Devin release notes for November 11 — Devin can set itself reminders to resume a task later, including in sessions launched by automations; design mode (Design Mode) supports pencil annotations in five colors, and a security profile option prompts before loading external content. 🔗 source
- Devin release notes for November 12 — Each tool on an MCP server can now be set to allowed, requiring approval or blocked (Allow, Needs approval, Block); the redesigned Devin Review page opens to all organizations, and Devin imports Veracode and Black Duck findings in closed beta. 🔗 source
Open models and research
- Naxi translation — Andrew Peters (anthropology, Harvard) fine-tunes Qwen3.5-4B-Base on about 50 000 lines for this language with roughly 300 000 speakers, using a dictionary of 3 619 entries. The gain over the model it replaces is not significant, and 276 of the 446 tested lines overlap with the training data, according to the author. CC BY-NC 4.0 license. 🔗 source
- Harel, at the Lexsi Labs hackathon — A verifier the agent cannot see validates only 4 of the 11 tasks (out of 24) on which it passes the visible tests; after one round of DPO, the pass rate on these tests rises from 45 to 55 % on eight held-out tasks, but validated responses remain at 12 %, according to the authors. 🔗 source
What this means
The lucidml post puts memory at the heart of a playable world model. Keeping the entire history at full resolution would require too much computation to remain playable, according to the author; with what he calls Summaryformer, older frames survive only through their summaries, the initial image serves as a permanent anchor, and the attention buffer holds 6 528 keys instead of 26 208, roughly four times fewer. The author remains cautious about the result: a checkpoint that has not converged, context that is still short, sessions stopped by default after 180 seconds, and one CUDA GPU per player.
The project also shows what a single author can publish: a model he says he trained from scratch, its weights under a permissive license and a playable demo, but no training code, which limits reproduction to inference. Two other community posts from November 15 show the same caution: the Naxi model’s author flags a contaminated test set himself, and the Harel authors specify that this is a single, untuned round of DPO. In all three cases, the figures come from the authors.
For coding agents, the updates give users more control over what the agent does and what it costs. Devin now lets users configure each MCP tool individually, Qwen Code’s Auto mode checks that risky actions have been explicitly authorized, and Codex CLI, still in preview, passes cost and usage limits to status lines provided by an MCP server. The Harel result is a reminder of what is at stake with these controls: passing tests does not prove that the code is correct.
Sources
- lucidml’s post about his GTA V world model (Hugging Face)
- lucidml/gta model (Hugging Face)
- Playable demo, lucidml/gta Space (Hugging Face)
- Codex CLI 0.171.0-alpha.6 (GitHub)
- Qwen Code v0.27.4 (GitHub)
- Devin release notes for November 11
- Devin release notes for November 12
- Andrew Peters’s post about the Naxi model (Hugging Face blog)
- Harel, Lexsi Labs hackathon (Hugging Face blog)