Dario's Open-Source Warning Sparks Controversy as GPT Astra Nears 70% Release Probability This Week

Dario's open-source warning draws pushback as model rumors and open-source moves intertwine, heating up security debate.
The November 17 AI industry updates reveal multiple tensions. Anthropic CEO Dario warned at a hearing that unrestricted open-source models could spiral out of control, drawing public criticism from VC Bilger that he's using the safety narrative to suppress open-source competition. On model releases, Polymarket puts GPT-6.2 Astra's probability of launching this week at about 70%, Zhipu's GLM5.5 is rumored to launch November 20 but lacks corroboration, and Google's model naming chaos drew complaints. Ahead of Ignite, Microsoft exposed two experimental code models including MyCodeR Flash, while OpenCode rolled out a $10-for-$100 GLM 5.3 usage deal (90% off). In the open-source ecosystem, Meituan open-sourced LongCat Video Avatar 1.6 under MIT, capable of generating 10+ minute videos, and Tencent open-sourced the self-hosted AI delivery platform LoopForge. Additionally, Apple adjusted macOS permissions to address AI Agents, Anthropic reports ~90% of its code is now autonomously written and verified by models and has poached a former DeepMind executive, and NATO is partnering with OpenAI to strengthen cybersecurity defense.
The November 17 AI evening roundup packs a lot of information—from model release rumors to public sparring among industry heavyweights and new moves in the open-source ecosystem, several threads intertwine to sketch out some real tensions in today's AI industry.
Dario Warns of Open-Source Risks, Investor Pushes Back Publicly
Anthropic CEO Dario Amodei issued a warning at a hearing: unrestricted open-source models, if allowed to keep developing, could pose extremely high risks. His core argument is that once a model leaves the developer's hands, there's no way to monitor or revoke unauthorized access.
That statement quickly drew pushback. Well-known venture capitalist Bilger reposted a critique, pointing directly at closed-source giants using lobbying power to block open source—essentially because open source constitutes the only real risk to their business strategy.

This debate isn't new, but the timing is telling. The capability frontier of open-source models is rapidly approaching closed-source flagships, and the fight over regulatory discourse has reached a fever pitch. Dario's concern has its technical logic—once model weights leak, they're indeed hard to recover; but Bilger's rebuttal is equally sharp: the safety narrative is sometimes also a competitive tool. Both sides have support for their positions, and consensus seems unlikely in the short term.
The core distinction between open-source and closed-source models lies in whether model weights are public. Open source (such as the Llama series, GLM series) releases trained parameter files that anyone can download, deploy locally, or even fine-tune; closed source (such as GPT, Claude) only provides APIs, with weights kept on the vendor's servers. What Dario means by "unable to monitor or revoke unauthorized access" is that once weights leak, vendors can neither see who's using them nor shut them down remotely. This does pose a security challenge, but it's also precisely why the open-source ecosystem innovates quickly and deploys at low cost. The regulatory debate typically centers on "capability thresholds"—how capable a model must be before openness should be restricted—and there is currently no international consensus on that threshold.
Model Release Wave: GPT Astra Betting Odds at 70%, GLM5.5 Rumors Swirl
On the prediction market Polymarket, the contract betting on GPT-6.2 Astra releasing this week has risen to about 70%, with trading volume around $186,000. The highest probability is assigned to a November 16–22 release, with only 16.8% betting on a delay to next week, and about 52% of contracts betting on a release tomorrow night.
On the other side, new word has emerged about Zhipu's next-generation flagship large model GLM5.5. A screenshot from an unofficial group shows a Volcano Engine marketing staffer revealing in a phone conversation that the model will officially launch on November 20. Most people are currently taking a wait-and-see attitude, and the leaker himself says he doesn't quite believe it, with no other corroborating GLM5.5 information for now.
Model naming is getting harder and harder to keep up with. The Gemini 4 Neo model reported this morning has now been removed from the list, and the poster complained that Google's model naming is impossible to remember every time a new name appears. The internally codenamed PaperGum Volcano Nano Banana Pro 2 has already been deployed to Google Cloud, but according to the tester who posted the message, this version performed rather poorly in his tests, and actual results may fall short of expectations.
Polymarket is a blockchain-based prediction market where users bet stablecoins on binary events (such as "whether a model will release in a given week"), and the contract price equals the market-implied probability. For example, a contract quoted at $0.70 means the market believes there's about a 70% chance the event occurs. Such markets are often used to reflect insider information and expectations from the tech world in advance, but due to limited liquidity and susceptibility to being moved by small amounts of capital, their probabilities should be treated as reference rather than confirmation. The "trading volume of $186,000" mentioned in the article refers to the cumulative trading scale of that contract—the larger the scale, the more informative the price usually is.
Microsoft's Dual Code Models Exposed, OpenCode Offers 90% Off GLM Usage
Ahead of Ignite, Microsoft appears to have prepared two new code models. The Foundry model card link for MyCodeR Flash is already live, and the GitHub Copilot catalog has also revealed specs for MAI CodeR and the Flash version. Both are labeled experimental previews, supporting 1.05Mbps context and 128K output, with the ability to input images and PDFs, targeting different subscription tiers respectively.

OpenCode is rolling out a limited-time perk for all OpenCode Go subscribers: spend $10 and get $100 worth of GLM 5.3 usage—a 90% discount. This campaign, called Operation GLM Phase 1, is available to everyone, but the official announcement doesn't specify an end date, so those who want it should claim it early.
Context window refers to the total number of tokens a model can read in a single pass, determining how much code or documentation it can process at once. The "1.05Mbps context" in the article should be an ultra-long context of about 1.05 million tokens (M meaning million-level), far exceeding the 8K and 32K levels of early models, enabling the model to swallow an entire medium-sized code repository in one go. Experimental preview means the model has not yet reached general availability (GA), its specs and interfaces may change at any time, stability is not guaranteed, and it's typically aimed at developers for early testing and feedback.
Open-Source Ecosystem: Meituan's Digital Human Model and Tencent's Delivery Platform
Meituan has open-sourced its next-generation audio-driven digital human model LongCat Video Avatar 1.6 on the ModelScope community under the MIT license, capable of extending continuously generated digital human videos to over 10 minutes. The new version optimizes lip-sync and body movements, reducing frame flicker and identity drift in long videos. It ranked second in an Arena human preference blind test with over 30,000 votes, and also offers INT8 quantized weights for easy deployment.

Tencent has open-sourced LoopForge, a self-hosted AI software delivery platform that can be deployed on private servers or in Docker sandboxes, orchestrating up to 10 mainstream programming tools including Cloud Code, Codex, and CodeBody. The platform offers two modes—conversational MainBot and standardized SOP workflows—splitting development into multiple division-of-labor nodes, with support for failure tracking, node re-runs, and template reuse.
Apple has adjusted full disk access and device control permissions in the macOS 27.3 Beta, splitting the previous quick authorization into fine-grained per-app authorization to cope with increasingly autonomous AI Agents. Although developers complain about more pop-ups, the industry believes this will make it safer and more worry-free for ordinary users running Agents. Beta2 also adds a panel to the menu bar for easily seeing which apps have bypassed protections.
Audio-driven talking avatar is a class of video generation models that automatically generates corresponding lip movements, expressions, and body movements based on input speech. Its difficulties lie in identity drift (the character's appearance gradually deforming as generated frames increase) and lip-sync (alignment of mouth shapes with phonemes) in long videos. The industry commonly uses human preference blind tests (such as Arena) to evaluate generation quality—users vote between two options without being told which model produced which, and results are tallied using Elo-style rankings, which is closer to subjective perception than automatic metrics.
Anthropic's Internal Practices and Talent Moves
The CloudCode team has achieved in internal development about 90% of code being written, verified, and fixed by the model itself, with developers only needing to define completion criteria, and the AI autonomously runs end-to-end tests and code reviews. To prevent the model from getting stuck in infinite loops, the system has an anti-deadlock mechanism that allows a third pass after two consecutive interceptions; unstable tests are simply deleted with the reason recorded. You can replicate this workflow by configuring Cloud.md and validation skills in your project.
At a moment when competition between Gemini and Cloud is white-hot, Anthropic has poached Jaz Seckon, former Chief Strategy Officer at DeepMind, who will work full-time as a cognitive rigor coach for Cloud, teaching AI to quantify its own ignorance.
Security Defense Cooperation
NATO and its allies' cybersecurity agencies are collaborating with OpenAI to use AI to help defenders find system vulnerabilities, test fixes, and protect critical infrastructure. So far, both sides have only disclosed that the partners are trusted defense agencies.

Looking at this news alongside Dario's open-source warning is quite interesting: on one hand, closed-source vendors emphasize runaway risks; on the other, a military alliance partners with an AI company to harden defenses. The security issue is evolving from a technical discussion into a multi-dimensional game of policy and business.
Related articles

Abnormal Error Rate Spike in Claude Mythos 5.2 and Fable 5.2
Claude Mythos 5.2 and Claude Fable 5.2 show an abnormal spike in request error rates, marked as investigating on the official status page. This article covers the incident overview, practical impact on developers, and recommended responses.

ReAct Runtime Context Trimming: Why Claude Code Acts at 70%
A deep dive into ReAct runtime context trimming: why Claude Code compresses at ~70% usage, why tool calls dominate, why messages must be dropped in pairs, and why codebases don't need RAG.

AI Agent vs LLM: How Cursor and Claude Code Actually Work
An in-depth analysis of the essential difference between AI Agents and LLMs: LLMs generate code suggestions, while Agents modify files and run tests via tool calls. Using a Spring Boot bug fix as an example, it breaks down the workflows, permission control, and verification mechanisms of Cursor and Claude Code.