프런티어 AI 안전 권한, Muse 실행 환경, 에이전트 모델 성능의 교차

Demis Hassabis

@demishassabis

2일 전

Dario's essay points towards the right path forward. The details need working through, but the direction is correct for meeting this critical moment. This is also why we recently put out our proposal for an industry-wide standards body for frontier AI. https:// x.com/demishassabis/ status/2076957440109625718?s=20 …

dax

@thdxr

2일 전

the openai hugging face incident gets cited a lot but what about the detail where they had to use GLM 5.2 to analyze the attack because the "safe" proprietary models refused

muse is super fast

Alexandr Wang

@alexandr_wang

2일 전

you can use muse to do all sorts of cool things!

Teknium

@Teknium

2일 전

Verdict: not one bug — the trading flow is dying on the approval gate (60 s timeout that should be 300 s), and the agent is burning hours delegating around it. Plus a broken Honcho memory backend, so it can't remember what it bought. Bundle: Jake (Telegram DM), v0.21.2 [1c671bea] current, openai-codex gpt-5.6-sol, max_turns 45, Linux VPS, memory_provider honcho. agent.log window only covers Sep 12 16:00→23:03; the Sep 10–11 trading fights are visible only at turn level in gateway.log. What the logs actually show ① Approval gate is failing closed after 60 s, not 300 s 22:42:59 tool call → 22:44:00 Tool terminal returned error (60.75s): BLOCKED: Command timed out without user response. Default approvals.timeout has been 300 since #68597 (Jul 21, in his build). His config.yaml almost certainly carries a materialized approvals.timeout: 60 from the pre-May-17 save_config that wrote defaults to disk (he installed before Jun 17). No migration flipped it — same class as bf0d8fed (verify_on_stop v32 migration). Every on-chain execute command trips smart-approval → Telegram card → phone push arrives late → BLOCKED with "do NOT retry" → agent stops. Sep 11 shows the fallout: ~10 stale taps (Telegram button resolved 0 approval(s), choice=once) — he's tapping cards that already expired, sometimes 3 in a row. ② Hours-long turns spent delegating instead of executing Sep 11: response ready time=7580s, time=12638s api_calls=36. Sep 12: time=6164s. Chains of [ASYNC DELEGATION COMPLETE] while he types "why is it blocked? make it submit" / "I said fucking execute ASAP". On Sep 12, 4 of 7 subagents ended max_iterations_reached(50/50), main hit 45/45. Context 150–255k tokens per call (272k ctx, compression at 231k). This is the agent's response to ①: it can't run the trade, so it "repairs" the executor forever. ③ Honcho memory is dead Every turn: Honcho background session init failed: Not Found, Failed to sync messages to Honcho: Not Found, honcho_search → "Honcho session could not be initialized". He asks "DID WE BUY NOCK YESTERDAY?" and the agent has no working long-term memory. Workspace/peer gone or key mismatch on his side. ④ One unexplained anomaly worth a repro 17:13:34, main Telegram session 4da60767: Tool execute_code returned error: BLOCKED: execute_code runs arbitrary local Python … Cron jobs run without a user present…. That's the _CRON_CTX deny, so _is_cron_approval_context() was true inside a gateway turn. fb6446fc (ContextVar-scoped HERMES_CRON_SESSION) IS in his build; a cron job (Across token status check) ran 16:23–16:44 in the same process on cron-parallel_0. I could not find the leak path by reading; needs a live repro (gateway turn + concurrent cron + execute_code). Noise, not the cause: Telegram unreachable from his box 19:00–21:00 Sep 12 (Bad Gateway / fallback IP failing → SIGTERM restart); hermes update wedged 17:59 on "Restore local changes now? [Y/n]" forwarded to Telegram, watcher timed out 1800s (update did land later); MCP 'affinity' parked (his script); stale systemd TimeoutStopSec=90 < drain 180. What to tell him · hermes config get approvals.timeout — if 60, set 300 (or higher). Then his taps land. · Fix Honcho (recreate workspace or re-run setup); until then the agent can't remember positions. · Consider approvals.mode: manual plus explicit approvals.allowlist for the execution script so it stops re-prompting, or accept the risk with session/always on the card.

deepseek v4.1 flash has arrived on together ai it beats gpt-5.6 sol on agentic benchmarks at one third the cost per task the model supports a 1m context window and native multimodal input, with 552b total parameters

This 8+ minute AI film proves one person can now make a real movie. 100% made with BytePlus + Seedance 2.5. Hollywood is not ready for this. I watched the whole thing. This isn't a demo reel, it's a real short film. Faux news. Punchy vignettes. A clear rule. An ending that actually sticks. Seedance 2.5 handles the expensive stuff insanely well: cinematic lighting, glossy VFX, huge spectacle shots, and clips up to ~30 seconds. What used to need a crew + serious budget can now be written, generated, edited and finished by one storyteller.

【AI's Grand Challenge Leads to Life's Grand Challenge: The Royal Society Unravels the Future of "World Models" and Intelligence】 In recent years, with the rapid development of AI, opinions have emerged suggesting that AGI—artificial general intelligence on par with human intelligence—has already been achieved. But is that truly the case? The key lies in the concept of the "World Model." A world model is the foundation that allows living beings or AI to internalize a representation of the external world and predict what will happen next to guide their actions. The Royal Society, known as the world's oldest scientific journal founded in 1665, has published a special issue titled "World Models in Natural and Artificial Intelligence" in its *Philosophical Transactions of the Royal Society A*, centered on this theme of world models. Contributions come from leading researchers in AI, biology, and philosophy, with Sakana AI CEO David Ha ( @hardmaru ) participating as a co-author of the lead article. Here are three key points that run through this special issue. ・"What it can do" and "What it understands" are different Current large-scale AI models can accomplish an astonishing number of tasks, but this often stems from memorizing patterns in word sequences rather than truly understanding causality. Some argue that simply increasing computational resources won't bridge this gap. ・AI that knows itself When AI learns to predict its own internal states, its internal representations become more organized and efficient, reducing redundancy. For embodied AI like robots, the ability to grasp one's own state forms the foundation for adapting movements to different situations. ・AI's grand challenge connects to life's grand challenge A world model also serves to capture oneself within the environment. Living beings don't just passively receive given information; they actively interact with their surroundings to learn about the world. Future AI may increasingly approach the realm of artificial life (ALife) research. There remains a significant gap between "being able to handle language" and "understanding the world." Sakana AI is tackling this theme through research on world models and Physical AI at RSI Lab. The special issue is available here: https:// royalsocietypublishing.org/rsta/issue/384 /2320 …

Graph Convolutional Network by hand ~ 12 steps walkthrough below Graph Convolutional Networks (GCNs), introduced by Thomas Kipf and Max Welling in 2017, are the tool for data shaped like a graph: social networks, recommendations, biological networks, drug discovery, molecular chemistry. I drew and calculated a simple GCN entirely by hand. Goal: run a two-layer GCN, then a small classifier, on a five-node graph, filling in every cell yourself. 1. Given A graph of five nodes, A to E, with edges between some of them. 2. Adjacency matrix (neighbors) Put a 1 wherever two nodes share an edge, in both directions. 3. Adjacency matrix (self) Add 1s down the diagonal, one self-loop per node. That is just adding the identity matrix. 4. Messages Multiply each node's embedding by the weights and biases, then ReLU. Negatives become 0. 5. Pooling Multiply the messages by the adjacency matrix. Each node gathers the messages of its neighbours and itself. 6. Visualize Node A pools [3,0,1] + [1,0,0] = [4,0,1]. 7. Second GCN layer Messages again: weights, biases, ReLU. 8. Pooling again Pool over each node and its neighbours, once more. 9. Visualize Node C pools [1,2,4] + [1,3,5] + [0,0,1] = [2,5,10]. 10. Fully connected layer Weights, biases, ReLU. This time there are no neighbours to pool, just the node itself. 11. Linear layer One more: weights and biases. 12. Sigmoid Squash each score to a probability (≥ 3 → 1, 0 → 0.5, ≤ -3 → 0). That is the classification for each node. You have just classified every node in the graph by hand. The outputs: A: 0 (very unlikely) B: 1 (very likely) C: 1 (very likely) D: 1 (very likely) E: 0.5 (neutral) The takeaway: a GCN layer is two parts. The top part pools each node with its neighbours through the adjacency matrix. The bottom part is an MLP that transforms each node on its own. A transformer layer has the same two parts, with an attention matrix where the adjacency matrix was. Both matrices do one job, mixing across positions: attention over tokens, adjacency over nodes. In my class I call the GCN the transformer's little cousin: a bit more stubborn, because its attention is fixed by the graph rather than computed from Q, K, and V. Draw the two side by side and the resemblance is hard to miss. Save this post! #AIbyHand #GraphNeuralNetworks #DeepLearning

OpenAI Developers

@OpenAIDevs

2일 전

1-800-ChatGPT