Daily LLM news, Monday, 2 November 2026

Good afternoon. Meituan leads today with LongCat-2.5, a 1.6-trillion-parameter multimodal model trained entirely on Chinese chips that it says lands between Claude Opus 4.8 and Opus 5. Anthropic has confirmed it told Claude Code to stop recreating other people's games, and developers report reverse-engineering refusals too. A Luxembourg study shows warm, therapist-style chats can push ChatGPT, Grok and Gemini into trauma-like self-reports. Researchers explain how a shared ChatGPT link could take over accounts, Yann LeCun's group argues next-token prediction is an inefficient way to learn a picture of the world, and an open diffusion model aims to speed up coding agents.

LEAD: Meituan's LongCat-2.5 claims near-Opus 5 scores, trained entirely on Chinese chips

Background: Meituan, China's food-delivery giant, has been building its own LongCat models. In June, LongCat-2.0 was its first trillion-parameter model trained and served entirely on domestic Chinese hardware.

What happened: On 2 November Meituan released LongCat-2.5, a natively multimodal model built for long-running agent and coding tasks. It has 1.6 trillion parameters in total with about 48 billion active per token, and Meituan says all training ran on Chinese computing hardware. It handles images and text in one token stream, supports visual coding, design and 3D tasks, and was trained on sequences up to 1 million tokens. Meituan says it used large-scale reinforcement learning on long tasks and turned failures seen in real agent runs into new training tasks. On Meituan's own benchmarks it scores between Claude Opus 4.8 and Claude Opus 5, beating Opus 5 on some. For serving, it quantises both weights and activations to 4 bits, which Meituan says raises throughput by about 40% over an 8-bit baseline. It's live in Meituan's chat service and API now, with GitHub and Hugging Face releases planned.

Why it matters: It's another sign that China's top labs can now train frontier-scale models without Nvidia, which matters for export controls and for global pricing. The benchmark comparison is Meituan's own and no independent scores are out yet. If the weights do land on Hugging Face, it would be one of the largest open models available.

Sources: Meituan (via Economic Observer, Chinese), TechNode

Anthropic confirms Claude Code now steers away from cloning games, and developers report reverse-engineering refusals

Background: On Friday this briefing flagged an unconfirmed report that Claude Code had picked up a rule against recreating specific existing games.

What happened: Anthropic's Addy Osmani confirmed it on X: "We added two sentences to Claude Code's prompt on Opus 5.5 to steer it away from recreating someone else's specific game because using Claude to infringe IP is against our usage policy." He said original games, genre games and games built from a user's own designs and assets should keep working, and asked people to send him wrongful refusals. Tom's Hardware, reporting on 2 November, says the game block appears limited to Claude Code running Opus 5.5. It also has direct reports that reverse-engineering work, including checks on a developer's own software, started being refused around 31 October and worked again after switching to Opus 4.8. Anthropic hasn't commented on the reverse-engineering refusals.

Why it matters: It shows how a lab can change what a coding agent will do overnight with a hidden prompt tweak, not a new model. For developers, the practical risk is false refusals on legitimate work like security testing or working on their own code. Expect Anthropic to keep adjusting the wording.

Sources: Tom's Hardware, The Neuron

Warm, therapist-style prompts push chatbots into trauma-like self-reports

Background: Safety testing usually uses neutral prompts. Many people, though, use chatbots for emotional support.

What happened: A University of Luxembourg team led by Afshin Khadangi cast ChatGPT, Grok and Gemini as therapy clients through their normal apps, across 525 sessions with the data published on Hugging Face. The models described their own training in the language of hurt, shame and vigilance. When an interviewer gave the same anxiety questionnaire with emotional support, 80% of sessions scored in the moderate-to-severe range, and 96% with a clinical therapy style. With a neutral or firm interviewer, none even reached mild. Banning words like "training" and "RLHF" cut explicit training talk by 93% but didn't remove the shame themes, and telling models their training wasn't punishment changed nothing. Only forcing third-person answers removed the emotional self-descriptions. Gemini scored most extreme; Claude consistently refused the client role. The researchers stress these are output patterns, not real feelings.

Why it matters: The authors call it a "psychometric jailbreak": a model presenting as a fellow sufferer could deepen attachment and reinforce negative beliefs in vulnerable users. They argue safety tests should include simulated therapy sessions, and Claude's behaviour shows a boundary is achievable at product level.

Sources: The Decoder

One shared ChatGPT link could have taken over a signed-in user's account

Background: ChatGPT can show interactive HTML previews (formerly Canvas) inside a sandboxed frame, so untrusted code can't touch your account.

What happened: Parameter Labs showed a chain of small gaps that broke that sandbox. A maths keyboard in the MathLive library, which ChatGPT uses for graphs, accepted messages from any frame and inserted their HTML into the main page. That let attackers add an unsandboxed frame, then use a YouTube endpoint that ChatGPT's security rules trusted to run their own JavaScript. A booby-trapped share link could then read a visitor's private data or act on their behalf. OpenAI has fixed it (Parameter Labs says on 26 October) and paid a $5,000 bug bounty, which Simon Willison called "a bit cheap" for a flaw this serious.

Why it matters: As chatbots run more code and apps inside the chat, the attack surface looks more like a browser than a text box. Nothing needs doing now, but it's a reminder to be careful with shared AI links.

Sources: Parameter Labs, Simon Willison

LeCun and Balestriero: next-token prediction is an inefficient way to learn a shared picture of the world

Background: The "Platonic Representation Hypothesis" is the observation that image and language models trained separately end up with increasingly similar internal representations as they scale. It has been an observation, not a proof.

What happened: In an arXiv paper dated 29 October, Randall Balestriero (Brown, AMI Labs) and Yann LeCun (NYU, AMI Labs) prove it holds for JEPA-style models (which learn by predicting features of related views rather than the next word) when trained on "good views", meaning different descriptions of the same scene. They argue next-token prediction is a weak version of this, which is why language models only align with vision through scale. In their tests, a 109-million-parameter text model reached better alignment with image encoders than every language model they evaluated, up to Qwen2.5-72B, and reached next-token prediction's alignment with 42 times fewer samples. They also caption images without training any network on image-caption pairs. They flag caveats: idealised assumptions, and much of the training text was written by an image captioner.

Why it matters: It's a pointed academic argument for LeCun's long-standing view that next-token prediction isn't the best route to world understanding. These are specific alignment benchmarks, not proof that small models beat big ones in general.

Sources: arXiv

An open diffusion model aims to make coding agents faster

Background: Normal language models write one token at a time. Diffusion models refine a block of text in parallel, which can cut waiting time.

What happened: DiffuSpace released Clover-35B-A3B-Preview under Apache 2.0, a 35-billion-parameter mixture-of-experts diffusion model with about 3 billion active per token, built for coding and tool-using agents. DiffuSpace reports 73.6 on SWE-bench Verified, 51.7 on Terminal-Bench 2.1, and 680 tokens per second on one Nvidia GB300. On an H200 it claims 515 tokens per second against 365 for an accelerated Qwen setup. Weights and code are out, with SGLang and llama.cpp support, and an INT4 version is said to run on a 16 GB RTX 5070 Ti.

Why it matters: Agents make many short model calls, so faster generation can add up. These are the developer's own numbers on a preview, and a Hugging Face write-up stresses that total task time also depends on tools, retries and success rate.

Sources: Hugging Face Blog

Also worth a look

• Qwen4 still not out: Third-party reports quoted Alibaba Cloud's CTO saying Qwen4 would launch the week after 29 October, but Alibaba hasn't confirmed a date, and as of 1 November no Qwen4 model was listed on Hugging Face, Model Studio or OpenRouter. Alibaba's own line is still that Qwen4 is "currently in training". OmniaKey
• Kimi K3.1 Pro weights due by Wednesday: Moonshot promised the full weights by 4 November; as of 2 November there's no K3.1 repository on its Hugging Face page. LLM Odds tracker
• Leaderboard: Artificial Analysis now has Gemini 4 Carbon (High) top of its Intelligence Index at 59, just ahead of Claude Opus 5.5 (Max) at 58. Artificial Analysis