# Dan's Radar — Week of November 2, 2026 ## Executive summary Apple opens **Siri AI** to all developers, letting any app expose actions and content through App Intents months before the customer rollout. Revolut reportedly saw AI gains only after leaving OpenAI and Anthropic for self-hosted GLM, Decagon runs 90% on **open weights**, and Anthropic's disclosure that Claude grades every chat for misuse pushed security users toward DeepSeek and Qwen. Google will reportedly cut web search for schools that opt out of AI, and Safari gets **vertical tabs** in the macOS 27.3 beta. On the device side, Cyclon (YC F26) pitched an **AI phone** to replace the iPhone, one investor called **Instinct** a textbook AI-boom failure, and robotics funding hit a record quarter. Opera passes the 100m active desktop users mark, making **local AI** models available to hundreds of millions of people in one go. ## Signals (23) ### 1. Apple opens Siri AI to all developers: any app can expose actions and content via App Intents, with skills that chain actions; ships to customers in months, App Store submissions in spring — 9.0 *Web · Apple Developer News · Tue* With the 27.3 betas, Apple's Developer Preview lets any app plug into Siri AI through App Intents: personal context, onscreen awareness, custom actions like booking or ordering, and new skills that combine actions into one experience. Siri AI can also surface any entity in the Spotlight index. Customer rollout in months; unavailable in the EU and for under-13s; server features have daily limits with paid expansion coming. > Dan's take: absolutely huge huge HUGE. investigating https://developer.apple.com/news/?id=gdk2umrf --- ### 2. OpenAI's autonomous pipeline found, exploited, and patched 948 Linux local privilege escalation bugs in one run for $1.2M in API credits — 7.5 *Web · OpenAI · Wed* OpenAI's pipeline, built on an upcoming cyber model, reproduced 1,724 Linux kernel vulnerabilities and produced 948 working root exploits plus candidate patches in a mid-October run costing $1,213,430, about $1,279 per exploit. It reports at least 10 novel exploitation techniques and over 5x more novel LPEs than a July run, and argues automated patching is now essential. > Dan's take: why do cybersec companies need to exist now? https://openai.com/index/a-thousand-fewer-paths-to-root/ --- ### 3. Uber: only 4 rides and 500 Uber Eats orders came through AI agents in its busiest week, out of ~320M trips and orders — 7.5 *X · TechCrunch (@TechCrunch) · Wed* On its Q3 call, Uber said that during Halloween week, its busiest of the quarter, four rides and about 500 Uber Eats orders were placed by AI agents against roughly 320 million trips and orders. DoorDash reported a few hundred agent orders. Khosrowshahi noted booking takes 50 seconds in-app, but said every AI company is talking to Uber. > Dan's take: this whole thing has been around mere seconds at the scale of large web cycles like CSS, Flash, HTML5 or React and virtual DOM rendering etc. not only will this execute in the following cycles, it will likely be a total sweep that will take the unprepared by storm. and in a transition period like this, shaping the web as you want it to be is better than trying to guess where the puck will be. underdog companies like Opera are in an incredible position to do this. https://x.com/techcrunch/status/2118100743987048581?s=46 --- ### 4. 'Instinct will be a textbook example of an AI boom failure': it rents distribution from messaging apps, so it owns neither UI nor compute, and Muse owns all three — 7.0 *X · Tristan Shin (@hooni_tri) · Sun* Tristan Shin argues a consumer agent needs distribution, UI, and compute. Instinct's messaging-app hack bought distribution but surrendered the interface, while Muse ships dynamic UI and owns the infrastructure. He finds continued Instinct funding after Muse baffling, and notes average consumers will not try ten agents. > Dan's take: there's good lessons here either way. https://x.com/hooni_tri/status/2117028061740949838?s=20 --- ### 5. Aside launches Computer Use: the YC browser agent now works outside the browser on Mac and Windows, claiming 2.7x faster with 65% fewer tokens than Codex on internal evals — 7.0 *LinkedIn · Aside · Fri* Aside, the YC-backed agentic browser that says 80% of work happens in the browser, adds Computer Use to cover the rest: it runs in the background across multiple apps on macOS and is available now on Mac and Windows. The 2.7x speed and 65% token claims versus Codex are from Aside's own internal evals. https://aside.com --- ### 6. Cyclon (YC F26) pitches an AI phone to replace the iPhone: ask it to do anything, it uses your apps, builds missing ones, runs tasks in the background — 6.5 *X · Adam Zinebi (@AdamZinebi1) · Mon* Cyclon, a YC F26 startup, says it is building an AI phone where anything you can do on the device you can ask it to do: the AI drives existing apps, generates ones you lack, and handles tasks in the background. 660K views in a day, no shipping details. > Dan's take: seems that the let downs that AI rabbit and AI Pin are far enough behind, we'll see more of these coming up https://x.com/AdamZinebi1/status/2117300841036353559?s=20 --- ### 7. Revolut reportedly saw AI gains only after moving from OpenAI and Anthropic to self-hosted GLM; Decagon runs ~90% on open weights — 6.5 *X · Oliver Molander (@OliverMolander) · Sun* Oliver Molander relays reports that Revolut's AI productivity gains came after switching to self-hosted GLM, and that Decagon runs about 90% of workloads on open-weight models, because cost and latency dominate mature agent workflows. His call: 80 to 90% of routine enterprise AI ends up on smaller open models. > Dan's take: local AI ftw https://x.com/OliverMolander/status/2116851762325757979?s=20 --- ### 8. Opera's desktop browser hits a record 100M monthly active users; Q3 preview shows revenue up 20% to $182.8M — 6.5 *X · 2logics (@2logics) · Tue* Opera reported 100 million desktop MAUs in October, up from 94 million in Q2 and 16% year over year, with preliminary Q3 revenue of $182.8M (up 20%), adjusted EBITDA of $43.2M at a 24% margin, and net income of $27.6M. Query revenue rose 17% and advertising 22%; full results come November 17. > Dan's take: Opera hits 100m desktop users while killing it on growth https://x.com/2logics/status/2117609137651499279?s=20 --- ### 9. Nous Research ships Arete-27B: Qwen3.8-27B post-trained on 53M tokens to cut 'party line' answers from 22% to 1%, 'the values in an open model are one fine-tune away' — 6.5 *X · Nous Research (@NousResearch) · Fri* Nous built an 803-prompt eval across seven value categories (free expression, democracy, privacy, and more) to measure how often a model adopts China's state line, then fine-tuned Qwen3.8-27B to bring that rate from 22% to 1% with no significant Hermes Bench loss. Eval, weights, and training data are public. 244K views. > Dan's take: the battle for influencing models has begun https://x.com/NousResearch/status/2118822925445050435 --- ### 10. Anthropic says Claude now grades every conversation for misuse right after answering; a critic calls it a safety report 'about you' and points to local models — 6.0 *X · Alexey Fateev (@superalesha) · Fri* Anthropic disclosed 'self-assessment', running since June: after each response Claude writes a preliminary safety assessment of the conversation, which it says has surfaced misuse other systems missed. Alexey Fateev reads it as a report filed about the user and says security people are moving to DeepSeek, Qwen, and GLM. > Dan's take: fascinating research https://x.com/superalesha/status/2116074374319018158?s=20 --- ### 11. NYT: Google will turn off web search for schools that opt out of AI for kids; DuckDuckGo: 'Forces AI on everyone, even school children' — 6.0 *X · DuckDuckGo (@DuckDuckGo) · Tue* Quoting a New York Times report that Google will disable web search for schools that opt out of its AI features for students, DuckDuckGo contrasts it with letting people choose whether AI appears in results. Google's position is not yet public; DuckDuckGo's framing is its own. > Dan's take: possibly a misscomm, but hey https://x.com/DuckDuckGo/status/2117660106976309521?s=20 --- ### 12. Safari gets a vertical tab option in the macOS 27.3 beta, with the top bar still in place — 5.5 *X · Andreas Storm (@avstorm) · Tue* Andreas Storm spotted a vertical tabs option in Safari in the macOS 27.3 beta. Early reaction notes Apple kept the horizontal top bar as well, so it is a sidebar added rather than the Arc-style layout users expected. > Dan's take: Safari, too, going vertical tabs https://x.com/avstorm/status/2117704238021177548?s=20 --- ### 13. Calif shows a PDF opened over AirDrop jailbreaking an iPhone in Lockdown Mode, with the compromise surviving the iOS 27.2 update — 5.5 *X · Calif (@calif_io) · Tue* Security group Calif posted a video of an iPhone 18 Pro with Lockdown Mode on being compromised through a user-opened PDF, then staying compromised after installing iOS 27.2. No technical write-up yet and no Apple comment; the persistence claim is unverified and the demo required user interaction. > Dan's take: the crazy thing here is the persistence https://x.com/calif_io/status/2117720348322144733?s=20 --- ### 14. Quest: a personal agent that 'studies you like a profiler', builds 100+ wiki files on you, runs its own Android phone, and manages Instinct and Muse for you — 5.5 *X · Vincent Nallatamby (@VNallatamby) · Thu* Quest asks for full access to your life in exchange for a memory-first personal advisor: 100-plus wiki files and a relationship map from day one, an Android phone it controls for WhatsApp, Instagram, and TikTok, generated iOS mini-apps per goal, and a proactivity loop trained on your feedback. 188K views on launch day. > Dan's take: LLMs are one directional and generic to billions, even if taking into account memory touch points etc. this reminds me of a Steve Jobs speech where he alluded at AI by giving the example of a student being able to interact with a virtual rendition of Aristotle thousands of years into the future, asking him questions about his motivation and chain of thought etc. many AI adjacent companies will try to build business cases around memory-powered AI. https://x.com/VNallatamby/status/2118388085256052967?s=20 --- ### 15. The Browser Company: biggest wave of Arc members into Dia yet; update adds Air Traffic Control and theme colors, top reply is 'BRO JUST BRING BACK ARC!' — 5.0 *X · Arc (@arcinternet) · Thu* Two weeks after the Arc sunset notice, the company says the largest cohort of Arc members so far has moved to Dia, and ships Arc's Air Traffic Control link-routing rules plus custom theme colors to ease the move. Replies lead with calls to bring Arc back; 8.9K views. > Dan's take: vox populi says "bro, just bring back Arc" https://x.com/arcinternet/status/2118410931193692494?s=20 --- ### 16. Dario Amodei: consumer AI video is 'existential' for OpenAI and Google, much of it is 'slop', and Anthropic will contract a model from someone else if it ever needs one — 5.0 *X · Veee (@vikktorrrre) · Fri* In a clipped interview, Amodei says Google and OpenAI are fighting it out in consumer because it is existential to both, that enterprise has no demand for 'cats riding donkeys', and that short-form AI video is largely fake, addictive slop Anthropic is not tripping over itself to enter. 76K views. > Dan's take: anthropic does not believe in microdramas https://x.com/vikktorrrre/status/2118763685564875020?s=20 --- ### 17. Vercel partners with SpaceXAI so Grok and Cursor can build, deploy, and scale on Vercel; 'it's happening' says the crowd betting on a SpaceX acquisition — 5.0 *X · Ally Bell (@withtenacity) · Thu* Guillermo Rauch announced a SpaceXAI and Vercel tie-up that lets Grok and the now SpaceX-owned Cursor build, deploy, and scale on Vercel. Ally Bell quoted it against an October 19 prediction that SpaceX will acquire Vercel. The partnership is confirmed; the acquisition is speculation. > Dan's take: not on my bingo card https://x.com/withtenacity/status/2118542800342208521?s=20 --- ### 18. AI video detection by absence: generated videos lack the near-inaudible low-frequency airplane Doppler sounds that real recordings carry — 4.5 *X · Pierre Beyssac (@pbeyssac) · Wed* Days after a method surfaced to geolocate videos from faint airplane Doppler noise plus FlightRadar tracks, Adam Holter reports that after Codex checked 120 generated videos, none contained those low-frequency Dopplers, making their absence a usable AI-video tell. 130K views. > Dan's take: this is insane. I'm preoccupied with AI content detection and this is a very clever way to do it. It makes me think of a time when we'll have to find this kind of creative angles when fighting a future rogue AI https://x.com/pbeyssac/status/2117974748281892884?s=20 --- ### 19. Robotics and physical-AI VC funding hit a record $18.6B quarter, per PitchBook — 4.0 *X · Joe Harris (@_joe_harris_) · Mon* A PitchBook chart shared by Alloy Robotics founder Joe Harris shows robotics and physical-AI venture deal value at a record $18.6B in the latest quarter, roughly triple a year earlier and the steepest bar on a 2021-to-2026 series. > Dan's take: one of the hottest fields in SV is robotics https://x.com/_joe_harris_/status/2117489704262062447?s=20 --- ### 20. Dia adds Link Routing and custom theme colors for the influx of Arc members; 72K views on the official post — 3.5 *X · Dia (@diabrowser) · Thu* Dia's own account of the same update the Arc account posted: Link Routing (Arc's Air Traffic Control) and customizable theme colors, framed as helping the past two weeks' wave of Arc members feel at home. The Dia post drew 72K views versus 9K for the Arc one. https://x.com/diabrowser/status/2118409797624053994 --- ### 21. Nordic's nRF71 chips: Wi-Fi 6, BLE, and Thread with 20x lower standby power, for smart-home and IoT devices lasting years on batteries — 3.0 *X · CNX Software (@cnxsoft) · Tue* Nordic Semiconductor's nRF71 series, starting with the dual-band nRF7120 and tri-band nRF7120E with an Axon NPU, combines Wi-Fi 6, Bluetooth LE, and Thread and claims up to 20 times lower standby power than the nRF70 line, enabling battery-powered Wi-Fi devices that last years. > Dan's take: so many possibilities https://x.com/cnxsoft/status/2117620699699536167?s=20 --- ### 22. macOS 27.3 renders Markdown files in QuickLook previews; Gruber: 'Starting to think Markdown might make it' — 3.0 *X · John Gruber (@gruber) · Wed* Guilherme Rambo spotted that the macOS 27.3 beta renders Markdown files in QuickLook previews, and John Gruber, who created Markdown in 2004, quipped that it might finally make it. Native OS support follows Markdown becoming the default text format for AI tools and agents. > Dan's take: I'm a fan of markdown. I do hope it makes it, and it looks like it https://x.com/gruber/status/2118125585272877161?s=20 --- ### 23. New York's Alza festival put vertical microdramas on an AMC screen; phone-first pacing and framing did not survive the theater — 2.5 *X · Hana Sachiko (@the_one_hana) · Mon* Hana Sachiko reports from the Alza festival in New York, where vertical microdramas screened in an AMC cinema to modest attendance; phone-first editing sat awkwardly on horizontal screens and the theater exposed pacing, framing, and sound flaws a mobile feed hides. https://x.com/the_one_hana/status/2117414740775346313?s=20 ---