Perplexity Open-Sources Decider v1.3, Google Ships Gemini 4 Hydrogen, and llama.cpp Reaches v1

Key Takeaways

1. Perplexity released Decider v1.3 and Decider Flash v1.3 under Apache licenses, with weights on Hugging Face and a free week on OpenRouter.
2. Google released Gemini 4 Hydrogen across AI Studio, the Gemini API, Search, and paid tiers, with partner-reported benchmark gains.
3. Hugging Face and ggml announced llama.cpp v1, the local model runtime's first fully stable major version.
4. SpaceX is building a supercomputing site near Brownsville, Texas, for frontier AI training, alongside water-infrastructure commitments.
5. Runway added Seedance 2.5 Flash for faster, cheaper video generation and said Praxis-1 now runs on SO-101 hardware.
6. Anthropic disclosed an SEC registration statement in connection with a proposed offering, alongside a livestream announcement.

1. Model Releases and Open Weights

1. Perplexity open-sourced Decider v1.3 and Decider Flash v1.3 under Apache licenses, publishing weights on Hugging Face and offering both free on OpenRouter for one week, through November 19. The v1.3 release adds a “locate” capability that finds target text from a description and returns character offsets, aimed at batch decision-making and evidence or entity extraction. The author claims significant gains over the best general extraction models on 17 out-of-domain NER, PII, and extractive QA tasks, and a 6.54 percentage-point accuracy improvement over the current classifier on 63 in-house production classification tasks. A cited post describes Decider reading 7,000 agent sessions totaling 1 billion tokens in one day, with the issues found leading to 14 agent-safety rule fixes — via 1 2
2. Google released Gemini 4 Hydrogen, described as targeting everyday coding, agentic tasks, and general use with a balance of speed and intelligence. It is available in Google AI Studio, the Gemini API, enterprise Gemini, the Gemini app, and Google Search, and to Google AI Pro and Ultra subscribers in Google Antigravity. Reposted claims include a 5-point accuracy edge over Gemini 3.8 Flash on the Box AI Complex Work Eval with under half the typical task time and about 12% fewer tokens, a 2.5x speed advantage over 3.8 Flash on certain tasks, and an Artificial Analysis Intelligence Index score above Argon; these figures come from reposted partner and community material rather than an independent evaluation. Google Search's Hydrogen upgrade was described as its largest model improvement to date and is rolling out globally — via 1 2 3 4
3. Hugging Face said Inkling 2 Small weights are available for download under Apache 2.0, crediting its team and safety-testing partners. Separately, Hugging Face and @ggerganov announced llama.cpp v1, calling it the first fully stable major version of the widely used local model runtime, running on phones, laptops, and workstations so users can run models locally without sending data off-device; the team committed to continued coverage of platforms and model types while keeping the project lean — via 1 2 3 4 5

2. Infrastructure and Enterprise Moves

1. SpaceX is building a supercomputing site near Brownsville, Texas, intended to train frontier AI for medical research, emergency response, advanced manufacturing, and national security. The first work is clearing waste illegally dumped on the site since the 1960s, including scrap metal, old appliances, rusted tanks, fiberglass insulation, construction materials, and thousands of illegally discarded tires; the site will generate its own power to avoid raising nearby residents' electricity bills. SpaceX has committed $220 million to expand Brownsville and Cameron County water and wastewater facilities, adding 20 million gallons per day of treatment capacity, described as the city's largest private water investment. The project has created more than 3,000 construction jobs with more than 500 long-term roles to follow, and related investment extends from the Rio Grande Valley to Cochise County, Arizona — via 1 2 3 4
2. Runway launched Seedance 2.5 Flash, which it says generates video faster and at lower cost. It also said Praxis-1 now runs on SO-101, built on thousands of datasets shared by the LeRobot community, and runs on a single RTX 5090 with NVFP4; checkpoints and installation instructions were provided — via 1 2
3. Anthropic announced a livestreamed fireside chat between JPMorgan CEO Jamie Dimon and Dario Amodei at noon ET. The same post carried a securities-law disclosure that Anthropic has filed a registration statement, including a prospectus, with the SEC for a related offering, and cautioned investors to review the prospectus and other SEC filings; no securities may be sold and no purchase offers accepted before the registration statement takes effect, and the post is not an offer to sell or a solicitation to buy — via 1

3. Security Tooling and Product Signals

1. Greg Brockman announced two cybersecurity products, Daybreak Defense Factory and Daybreak Red Team, saying both make frontier cyber capabilities easier to put to use and help enterprises defend in new ways. A cited post said the tools draw on experience defending OpenAI itself, where the team used frontier models to find and fix important vulnerabilities in its own code and infrastructure and is now turning that experience into tools and playbooks for other security teams — via 1
2. Perplexity product signals included continued growth in the Mac app's hybrid “cloud orchestrator plus local sub-agents” mode, with an invitation to try it; Producer generating a high-quality explanatory video on GPU internal inference, which a cited post said it produced; a cited SaaS customer said to have stopped using a SaaS tool because Perplexity Computer performed 3x better on some tasks, which the author commented on only as “interesting”; and a video-production Teammate said to have scheduled 48 sandboxes in parallel at 80 seconds per frame to finish a 60fps, motion-blurred photorealistic Pantheon video in 3 hours. These are author-shared product and customer anecdotes without independent verification — via 1 2 3 4
3. Google's smart glasses opened for preorder through a partnership among Google, Samsung, Gentle Monster, and Warby Parker, featuring a hands-free Gemini assistant, shipping December 10 with in-store demos at select locations; Demis Hassabis called smart glasses one of his favorite Gemini integrations. A reposted item also listed a dense Google release window from November 9 to 11 covering TPU 8tx and 8i max, Gemini Robotics 2.5, AlphaProof Code, Gemini 4 Carbon, Gemini Decision 1, and AlphaGenome 2, and speculated about Hydrogen or an early Krypton release — via 1 2

4. Practitioner Debate and Community Signals

1. At the AIE Code Summit, @lydiahallie's talk on Claude Code best practices drew 200,000 views in 13 hours, the highest first-day view count in AIE history, and the @aiDotEngineer YouTube channel reached 1 million subscribers at close. Closing talks framed the moment as a software industrial revolution: @trq212 argued the next step should be a “Renaissance” in which models work the way people do, enabling new languages, databases, operating systems, and interfaces and reviving decades of HCI research, but said that will not arrive automatically. @mitsuhiko proposed a “software workshop” over a software factory — fewer, better, more ambitious projects — noting he started 141 projects in a year while regularly maintaining only 10–15, and that agent task durations rose from minutes to over 40 minutes while his Pi coding agent's core tools barely changed. @dexhorthy described software engineering becoming “factory engineering,” treating coding agents as an engineering problem defined in code and measured on one's own codebase — via 1 2 3 4 5 6
2. Other practitioner notes included @RLanceMartin on multi-agent memory and “dreaming,” arguing memory should be selective and that stronger models should save strategic insight rather than everything or failure cases; @sqs predicting companies with code review but no PRs, departments, or roadmaps; a reposted claim that NVIDIA's commitment to open source is unchanged and that open source lets open and closed innovation ecosystems coexist; and a claim that “vibe-coding Salesforce” went from mocked to accepted because most users need only about 1% of Salesforce and can build that part with agents — via 1 2 3 4 5
3. Ethan Mollick said METR research conducted at Anthropic validated “vibe-testing,” matching his experience of Anthropic model quality and the AI consensus on X, and called NewSonnet 3.7 the most clearly above-expectations model of 2025. His generative AI lab released Primer, an agent-driven system that turns a teacher's lesson into a playable, pedagogy-oriented simulation for students, with the paper, prompts, and code all open. He also commented on a Culture-series visualization built from his Portals project repository, noting only one of the original 466 Portals was a Banks world and that it lacked types such as GSV — via 1 2 3
4. Kai-Fu Lee discussed AI valuation surges, Chinese open models creating opportunities for global developers, and enterprise AI transformation in a Bloomberg conversation, saying leaders often ask “I understand, but where do I start,” which he says his new book AI Native addresses. A reposted reaction to his forecast on AI-driven job losses by the end of the decade did not include the specific figure — via 1 2
5. Hamel Husain amplified evaluation-focused material, reposting a YouTube video on how a context-layer company does evals, described as a benchmark for enterprise retrieval and a best example of public evaluation, and reposting praise for Evals for AI Engineers, co-authored with sh_reya, alongside the advice that “look at your data” is the most useful current guidance in AI — via 1 2 3