大爱 AI 日报 AI Daily 2026/11/4: Anthropic Releases Cryptographic Proof Technology for Verifiable Reasoning Provenance

Today’s Summary

Anthropic releases FlashProve cryptographic verification technology, and Tencent Hunyuan Hyra 2.0 lets agents autonomously build scientific research tools such as protein design models.
From verifying model identity to cross-app collaboration to local writing assistance, everyone is embedding AI capabilities into critical nodes of concrete toolchains.
Developers should prioritize FlashProve and colibri today — the former affects security architecture, the latter determines the hardware bar for running large models locally.

🔥 Today’s Top 10 Highlights

1. Anthropic Releases Cryptographic Proof Technology for Verifiable Reasoning Provenance

Reasoning provenance can now be mathematically proven. An official Anthropic tweet announced that FlashProve is based on mature cryptographic techniques , enabling developers to verify whether a given response truly came from a specified AI model. The research claims its efficiency significantly outperforms existing open-source methods. These proofs can be used to prevent attackers from hijacking conversations, and also allow security evaluators to verify model identity.

2. Tencent Hunyuan Hyra 2.0 Lets AI Agents Autonomously Develop Scientific Research Tools

AI can now train models and fix chip defects on its own. A report from the Tencent team on Hyra 2.0’s scientific research system describes a three-layer agent collaboration framework called Harness, which distills debugging experience into a reusable, standardized skill library. According to the team, tools already produced include HyInv, a protein design model with a generation speed 57 times faster than commonly used industry tools, and HyProver, a mathematical proof workflow that the team claims has completely proven Yau’s circle action conjecture. The project is not currently open source.

3. Google’s Hard Bundle: Schools That Ban AI Will Also Lose Student Search Access

Schools that reject AI may also lose student search access. The New York Times reports on Google’s new arrangement for education accounts : starting in early 2027, AI tools will be off by default for K–12 student accounts, but schools that choose to fully restrict AI will simultaneously lose Google Search access for student accounts. Schools that allow AI can choose between two modes: a consumer version that retains AI overviews, or a Gemini student-guided learning version. The Los Angeles school district had previously hoped to keep search while disabling AI summaries; under the arrangement described in the report, schools would struggle to preserve both options at once.

4. OpenAI Builds a Voice Codex Prototype in One Week Using GPT-6 Astra

AI led the full design chain from hardware shell to software. According to a report on OpenAI building a voice Codex prototype in one week using GPT-6 Astra , two team members collaborated with GPT-6 Astra to complete the design of a Codex prototype device supporting voice interaction in one week, covering the enclosure, circuit board layout, interface, and software development. Actual manufacturing still requires multiple rounds of iterative review; this is currently a prototype demonstration stage.

5. iFlytek Spark X2.6 Hands-On: 293B MoE Architecture, 1M Context Window

A top-spec model trained on domestic compute has arrived. A hands-on article by Juejin author Silent King II shows that Spark X2.6 is based on a Mixture-of-Experts architecture with 293B total parameters, 30B activated per token, and a context window of 1M, targeting long-horizon complex tasks, code development, and agent execution. The author used it to complete a full embedded project for an ESP32 mini-piano workbench, stating that the model was trained entirely on domestic compute. This is a personal review, not an independent third-party benchmark.

6. Apple App Intents Lets Third-Party App Features Be Accessed by Siri AI

Siri can now complete tasks across apps. A post on Jike by Guicang relayed this Apple update : starting with the iOS 27.3 beta, developers can expose their app’s features to Siri AI for invocation, with a built-in Skill format that lets Siri determine capability boundaries. Cross-app scenarios include: asking Siri to change a pickup location after booking a ride, or sending an item saved in a shopping app to a friend in a messaging app. Previously, similar cross-app operation permissions for Doubao on mobile were widely refused by major domestic super-apps; Siri’s openness will likely bring similar tensions.

7. Open Typora 0.22.0 Adds AI Cross-Document Review to Local Markdown

A local writing tool has gained AI editing capabilities with per-change confirmation. Developer AniChikage posted an update in a V2EX thread on Open Typora 0.22.0: new features include selection-based rewriting with diff preview, batch modification plans spanning up to 4 documents, and local directory Q&A (with answers insertable as Markdown with footnotes). Numbers, dates, links, and code are protected by default; recovery points are automatically saved before any AI modification, and can be reviewed across batches after a restart. Well-suited for technical writers who need to maintain consistent formatting and terminology.

8. RF-Sensing Glasses Convert Silent Lip Movements to Text in Real Time

Mouth your words silently, and the glasses can recognize what you’re saying. The Subvocal Company (YC F26) presented their smart glasses prototype in a Hacker News Show HN post : antennas in the frame emit radio frequency signals that capture tongue, lip, and soft palate movements, which a model then decodes into text. In the demo, new users completed only 20 sentences of personal calibration, and cross-user facial structure variation remains the primary challenge. The team plans to deliver the first batch of 500 units in Q1 2027 and is currently accepting pre-orders; the project is in R&D and pre-sale stage.

9. The Gap Between Domestic Models and US Frontier Models May Become Visibly Wider

Compute constraints are the focus of this discussion. Diiiii on Jike relayed a summary of a Chinese LLM discussion , with two core judgments: the coding scenario has largely been mined out, and the next comparable-scale deployment scenario is not yet clear; domestic models face distillation difficulty, data bottlenecks, and insufficient compute, while overseas competitors continue to get stronger and cheaper. The article also references Anthropic’s Q3 monthly revenue growth trend to question whether the industry is at the peak of inflated expectations on the Gartner curve. This is a compilation of industry observers’ views, not official data.

10. Hong Kong AI Research Internship Openings in RSI and Inference Acceleration

Recruiting research interns in RSI and Personal Agent tracks. Jike user Five Strips of Rice Cake posted this Hong Kong AI research institute hiring notice , covering language model inference acceleration and recursive self-improvement (RSI). The position offers competitive pay, ample GPU and token budget, and interns can independently lead projects and be first author on papers. The affiliated institution has not been publicly named; interested PhD students in Hong Kong and those with relevant project experience can reach out through the original post.

◎ Industry Changes and Personal Impact

Ubuntu Has Not Delivered an Official Zero-Day Fix Kernel Publicly for Three Weeks

The RCE zero-day patch is still in the testing repository and has not entered the regular update channel. A Freenode article documents the fix progress for CVE-2026-98432 : Ubuntu’s official statement from November 2 shows that fixed kernels for 22.04, 24.04, and 26.04 LTS have entered the -proposed testing repository and are being validated on over two thousand certified servers and devices before moving to the regular update channel. 26.10 has already been fixed in 7.3.0-10.10; older supported versions are still pending. The vulnerability is officially confirmed to allow remote code execution and is not limited to MPTCP paths — disabling MPTCP is not a complete mitigation. Ops and server administrators can now manually enable the -proposed repository to test the patch early, or wait for it to be pushed through the stable channel.

⌘ Top Open Source Projects

JustVugg/colibri: Run Frontier MoE Large Models on Consumer Hardware

colibri’s tiered inference engine has published its core code , implemented in C with a three-level cache scheduler: hot experts reside on the GPU (supporting CUDA and Vulkan), warm experts reside in memory, and cold experts are streamed from NVMe, reducing the model’s dependency on VRAM capacity. Added 802 stars today, with a total of 47,247. Developers who want to run MoE models locally on existing hardware should first check the project’s supported models and hardware requirements before deciding whether to try it.

Usagi-org/ai-goofish-monitor: AI-Powered Intelligent Monitoring of Xianyu Second-Hand Listings

The ai-goofish-monitor project page shows this is a Playwright and AI-based multi-task monitoring and intelligent analysis system for Xianyu, supporting real-time and scheduled scraping with a full admin UI. Added 228 stars today, with a total of 16,365. Suitable for individual users and small sellers who need to watch for specific second-hand items or compare prices in bulk; Python stack, with customizable monitoring rules by condition.

◉ Social Media Picks

AI Coding Assistant Self-Retrospective Effectiveness: Only 2 Out of 7 Tests Actually Worked

Having AI revise its own working manual after finishing a task is far less impressive in practice than it sounds. Baoyu’s tweet cited paper research results from Singapore Management University and Heidelberg University : out of 7 valid tests, only 2 changes completely eliminated the target problem, and some changes caused the AI to perform unnecessary checks on unrelated tasks. Baoyu himself admitted he has always been skeptical of this kind of “self-evolving Skill” — without clear evaluation criteria, AI struggles to self-optimize effectively; relying entirely on human review is also costly. Developers currently using /retro or similar self-iteration Skills can reference this data to lower their expectations.

😄 AI Fun

Loaded a New Chess Engine onto an Old Computer — Won 78 Games, Lost 5

Nicholas Carlini took a 1996 Pentium Pro and loaded it not with period-appropriate software, but with Stockfish 17 rewritten as a 32-bit binary. The opponent was the same-era Crafty 11.13, running on a modern machine. Old Pentium vs. a modern machine running an old engine — the scorecard after 100 games : 78 wins, 17 draws, 5 losses. The hardware hasn’t moved in thirty years; it was the algorithm that won. If you’re using an old computer as an excuse to upgrade, ask yourself whether the software has given up yet.

❓ Related Questions

I Want Codex to Help Me Build a Voice Interaction Prototype — Where Should I Start Practically?

A good starting point is describing the interface by voice and having Codex generate a clickable draft. The OpenAI prototype experiment showed AI handling the enclosure and interface layout, with manufacturing still requiring human iteration. For those who want to keep advancing this kind of draft in ChatGPT, existing account holders can renew via 大爱·AI账号店’s ChatGPT Plus monthly renewal card .