Catch today's product, model, and platform updates in a few minutes.
A new study in Science reports that Google DeepMind's AlphaProof Nexus independently solved 9 of 353 open Erdős problems, including 2 that had stood for 56 years.
Kudos 3.0 has launched on Product Hunt, positioned as an AI wallet that helps you recover missed rewards. The browser extension analyzes your card wallet while you shop online and recommends the best credit card to use, with the company claiming up to $750 in extra earnings per year. Basic features are free forever, with optional paid Premium features available.
Tibo announced that the default context length for GPT-6.1 Sol in Codex and Work has been raised from 272k to 500k, an increase of about 80%. This reduces the occurrence of compaction and helps tasks that need to retain large amounts of context and detail over long periods; users can still override the setting themselves, and the model's maximum context length is 1M.
TechCrunch Disrupt 2026 kicks off in less than 48 hours. The three-day event runs October 13–15 at Moscone West in San Francisco, bringing together more than 10,000 founders, investors, and other industry attendees, with main-stage speakers spanning AI, robotics, transportation, and more.
Several executives surveyed believe voice AI has yet to see a transformative breakthrough comparable to ChatGPT. Investors have poured billions of dollars into voice AI startups, but technical challenges such as context understanding remain limiting factors, making it difficult for voice AI systems to run reliably end to end.
Microsoft CEO Satya Nadella posted on X on October 10 that AI's trust architecture should be reexamined: advanced AI systems need built-in isolation, independent control mechanisms, and an "emergency brake" allowing authorized personnel to pause or shut down models mid-run. He also called for isolating AI models from systems such as permission granting and operation logging.
Apple disclosed in a European Commission filing that it has reached an agreement with AI podcasting startup Huxe, hiring some of its team members and obtaining a non-exclusive license to its technology. Huxe was founded by former developers who worked on Google NotebookLM and shut down this May; the deal is an acquihire rather than a full acquisition.
TechCrunch Disrupt 2026 will take place October 13–15 at Moscone West in San Francisco, with over 300 startups presenting to 10,000 tech leaders, plus more than 250 speakers and 200+ sessions. Registering before opening saves up to $100, with a 50% discount on a second ticket.
TechCrunch compiled a list of AI agents you can talk to directly via text message, covering general assistants as well as scenarios for home, travel, and work. Users no longer need to download new apps—just state your needs as if texting an ordinary person, and these agents can remember context, connect to your existing apps and services, and complete tasks on your behalf.
Agibot is a personal AI agent that lives on your Mac, running in the background even after you close the window. It can send morning briefings, monitor changes, and handle web tasks on your behalf, asking first before risky actions. You can chat with it in-app, via Telegram or Slack, and it supports using your own API key.
Lune has launched on Product Hunt as a search engine built for research AI agents. Once connected, agents can search papers from top conferences, view abstracts, metadata, citation contexts and available full texts, track citations, and compare papers — supporting literature reviews, experiment design, writing, and peer review.
Maildun for Mac is live on Product Hunt: an email builder for macOS that lets you design by drag and drop, or plug in Claude, ChatGPT, Gemini, or Grok with your own API keys—AI suggests changes as proposals that you accept or reject. With the built-in MCP server enabled, Claude Code, Codex, or Cursor can also create and edit emails.
A study finds that while AI coding agents generate more code, this has not translated into more software actually in use, with efficiency gains absorbed by the "bottleneck" of human review. IBM statistics show that in 2025, 82% of developers use AI coding tools weekly, 59% rely on three or more assistants at the same time, and 78% report productivity gains.
Sierra has released a draft of the Personal Agent Protocol (Poppy) and announced new design partners. The open standard lets personal agents discover businesses, sign in users, and complete tasks via websites, APIs, and enterprise agents, with multiple companies already involved in the design.
The Nikon Small World in Motion microscopy video competition winner was disqualified for using generative AI, and the winner personally admitted to using AI in post-processing. Nguyen Nam Nhat from Vietnam has become the new champion with a video of a nematode and the single-celled Dileptus.
After integrating OpenAI's Daybreak AI agent, Sophos cut its average security incident response time from about 38 minutes to 89 seconds, reduced threat investigation time by 96%, and automated 52% of MDR cases end to end, while keeping human analysts in control of critical actions.
Open-source desktop AI agent OpenPilot is MIT-licensed and can connect to compatible models such as OpenAI, OpenRouter, Ollama, and LM Studio. It can read and write files in selected folders and run commands, with sensitive operations requiring user approval. A Windows version is currently available, while the macOS version is coming soon and not yet downloadable.
LegalOn Technologies selects between GPT-6 Luna, GPT-6.1 Sol and GPT-6 Astra based on task complexity, cutting Codex's estimated daily cost by about 65% versus GPT-5.5 while keeping development speed unchanged. The company also restricts Fast mode by default and sets monthly budget caps by department, team and individual.
Engineer Bo Lau has built an interactive website simulating Elizabeth Holmes' desktop from the Theranos case, based on over a thousand emails, text messages, slides, and documents from United States v. Elizabeth Holmes. Users can browse these real trial exhibits on the site, essentially a searchable archive of courtroom evidence.
OpenVids is an open-source, agent-first desktop video editing app for macOS and Windows. Users describe the desired effect in conversation, and multiple specialized AI agents complete the editing, add motion and sound effects, and check the render on a non-destructive timeline. Media stays local, under the MIT license.