Gemini 4 Argon Tops Vals Index as Flow Raises $50M and Runway Opens Its World-Model Stack

Key Takeaways

1. Google DeepMind released Gemini 4 Argon, which debuted at 68.9% on the Vals Index and leads multiple arenas, with Google citing agent-driven infrastructure work including 300+ TiB of memory freed in data centers.
2. Flow raised a $50M Series B at a $750M valuation, with backers citing adoption by Anduril, Joby, Stoke Space, Rivian, GM PPU and RV Tech.
3. Runway opened Runway Ads to select enterprise partners after internal data showed 1000% scaling, doubled ROAS and 34% higher conversion, and previewed the open-weight Praxis-1 world-action model for robotics.
4. Perplexity opened a no-account email entry point for Computer and open-sourced a context embedding model it says tops turbopuffer's context-bench.
5. Hugging Face released 200+ WebGPU kernels and a context embedding training method, while TRL hit 1M monthly training runs.
6. NVIDIA detailed VSS Blueprint 3.3 for manufacturing vision agents and framed Cosmos as a world foundation model for physical AI.
7. OpenAI announced an ASBDC partnership for small-business AI training, and a White House "superintelligence" agreement was signed by cross-industry leaders.

1. Frontier Models and Agent Infrastructure

1. Google DeepMind released Gemini 4 Argon, which it says improves substantially on key capabilities and is being rolled out first to governments and trusted cyber defenders through the Fairwind Program before broader availability. It debuted at 68.9% on the Vals Index, ranked 1st on Text Arena at 1525 and 8th on Code Arena: WebDev at 1679, with a blended cost of $8/MToken; it placed 1st in Coding, Hard Prompts, Instruction Following, Longer Query, Creative Writing and all evaluated occupational domains, and 20 points above the second-place Claude Opus 4.6 (High). On Artificial Analysis it showed a 15% hallucination rate versus 29% for Grok 4.7, 45% for GPT-6 Astra, 59% for Opus 5.5 and 69% for Fable 5.1, with a 50% correct-answer rate versus Opus 5.5's 66%, but it states when it does not know rather than fabricating. Google said its agents have freed over 300 TiB of memory in data centers, support a 1M-token output limit, and have been used to analyze performance data, find memory optimizations and apply changes at scale; they also contributed to migrating over 800,000 lines of kernel code from C/C++ to Rust and replaced 32,000 lines of SIMD code in a video decoder with safe Rust, yielding a 2.7x speedup over the existing Rust port with identical video output. — via 1 2 3 4 5
2. Perplexity opened an email entry point for Computer: tasks can be sent, forwarded or CC'd to [email protected] without a Perplexity account, free for a limited time, with each email task running as a regular Computer session viewable on web and mobile and retaining the same audit trail as in-app tasks. Perplexity also open-sourced its context embedding model, saying it performs best on turbopuffer's context-bench; the cited model, pplx-embed-v2-context-9b-preview, is trained so each document chunk is encoded with the full document as context and refreshes SOTA on ConTEB and turbopuffer's new non-public context-bench. A third-party evaluator said the model topped Perplexity's internal context-bench with document recall@10 more than 50% above traditional SOTA embedding models, and described iterating with the Perplexity team over several weeks with at least nine benchmark reworks, without privileged benchmark access. — via 1 2 3 4
3. OpenAI said small teams are using AI for more work across customer acquisition, product development and financial management, and published a new report on how small businesses put AI agents into practice; it also announced a new partnership with ASBDC to provide hands-on AI training and local guidance to more business owners. Separately, a White House "superintelligence" agreement was signed by cross-industry leaders, establishing that companies developing the technology bear primary responsibility for safe development and deployment, with accountability for falling short, reinforced by internal controls, independent external evaluation and board oversight; a reposted view said existing laws on cyberattacks, privacy violations, fraud and product liability already apply to superintelligence, and the agreement further requires transparency and audits. — via 1 2 3

2. Funding, Hardware Engineering and Robotics

1. Flow raised a $50M Series B at a $750M valuation, co-led by Antonio Gracias (Valor) and Gavin Baker (Atreides), with Sequoia Capital and Roelof Botha (SpaceX, Block) among the participants. A repost said Flow has become the default platform for requirements and validation at frontier hardware teams including Anduril, Joby, Stoke Space and Rivian, and that GM PPU and RV Tech, the Volkswagen-Rivian joint venture, are using its AI to reshape core development processes. The same repost argued that roughly 90% of mechanical engineers' daily execution work — CAD, Excel, simulation — is digital manual labor that AI agents will take over, shifting humans toward invention, architecture, trade-offs and judgment, a role change it called possibly the most profound since the invention of CAD, with AI moving from integration work to actual execution over the next year. The author compared Flow's effect on hardware engineering to Git+GitHub for software engineering, saying it aligns thousands of stakeholders and accelerates complex, irreversible, high-value processes from cars to rockets, and disclosed being a friend first and small investor second. — via 1 2
2. Runway opened Runway Ads, an autonomous engine for performance marketing, to select enterprise partners with broader rollout in coming weeks, disclosing internal usage since July: ad volume scaled 1000%, return on ad spend doubled, conversion rose 34% with click-through rate flat, cost per subscriber fell 41%, and it added $100M in ARR. Runway also released Praxis-1, an open-weight world-action model that applies video pretraining to real-world robot control, positioned as a policy model across arbitrary embodiments and environments for robotics developers and researchers; it is being tested with early partners including Noble Machines, Standard Bots and Ultra, with weights planned for release in coming months. Runway CTO Kamil Sindi introduced Solaris, described as the first world-interface model and a new kind of operating system that builds itself as it is interacted with, generating every frame in real time from user intent, toward a future where all software is generated without a line of code; co-founder and co-CEO Anastasis Germanidis said in the AI Summit opening keynote that general world simulators will be the most important technological development of our era. Runway Labs also introduced Project Continuum, a research app around real-time video interfaces, showing four new ways to interact with computers: Portals, Visual Thinking, Responsive Video Interfaces and Interactive Worlds. — via 1 2 3 4 5 6 7
3. NVIDIA detailed VSS Blueprint 3.3, which adds a Build Vision AI skill that the company says can build and deploy a vision AI agent for manufacturing lines in 30 minutes from a single prompt, with alerting, video search and incident reporting. NVIDIA also framed Cosmos as a world foundation model for physical AI: the term was coined by Jensen Huang to merge world models and foundation models, with "world model" used broadly to mean models that capture the world knowledge needed for a task — objects, environments, physical behavior — rather than the narrower robotics definition. A reposted view argued that different tasks need different world models but they describe the same physical world, so diverse data from specialized models can be integrated to train a single model for cross-task learning; a world foundation model is thus a "foundation model of world models," learning shared physical-world knowledge adaptable to many tasks and environments to help robots and autonomous systems anticipate consequences, evaluate possible futures and improve decisions. NVIDIA separately described AI infrastructure "fungibility" as the ability to flexibly run different workloads — all kinds of AI, all stages of AI, even non-AI workloads — on a platform built to maximize "productivity," remain "durable" years after deployment, and be "fungible" across every model and workload. — via 1 2 3 4

3. Open Source, Tooling and Developer Ecosystem

1. Hugging Face open-sourced 200+ WebGPU kernels for machine learning operations, calling them the fastest globally, optimized for specific hardware, running entirely locally in the browser and available through @huggingface/kernels. It also released a new context embedding training method in which each document chunk sees the whole document when encoded; pplx-embed-v2-context-9b-preview set new best results on ConTEB and turbopuffer's non-public context-bench. Telemetry showed TRL now reaches 1M training runs per month, against a goal of 1M per week, with each post-training run pushing AI ecosystem decentralization. A multi-harness RL setup added a capture proxy in openenv between harness and model, recording token ids and logprobs for every call while appearing to the harness as just another model provider; harbor provides tasks and sandboxes and trl trains with asynchronous GRPO. The same LFM2.5-2.6B weights solved 62% of held-out tasks in mini-swe-agent versus 33% in Claude Code; after RL across four harnesses the average rose from 42% to 54%, Claude Code rose from 33% to 49%, and tool calls fell 31%. Cloudflare Workers AI released its first trained models, clef, two fast and accurate decision models that it says lead quality and latency benchmarks, hostable on Workers AI or available as open weights from Hugging Face. — via 1 2 3 4 5 6
2. Hugging Face and Unsloth are co-hosting an open-source event with UnslothAI swag, demos of upcoming features, a DJ and food, entry code NOSLOTHSHERE. eidon_ai shut down and open-sourced its data: 9TB of first-person video, including 1,274 hours of household video, 7-point IMU arm tracking, and 13,451 records from 27 homes, under CC-BY-4.0, published on Hugging Face. A user said they trained their own Jev-style API from a single HuggingChat prompt, with ML Intern training on an NVIDIA A100 for 16 minutes and fixing its own bugs; the full 60K version cost about $6.60 and chose the right tool in 62% of roughly 20 scenarios. Hugging Face also launched an MCP data source into ML Intern, running on private open-source models, which it says makes it much easier for enterprises to fine-tune models on their own proprietary business data. — via 1 2 3 4
3. Sam Altman said DevDay feedback was positive with strong developer enthusiasm, including someone building a complete startup project in a day, and that 6.1 Sol is the fastest-growing model yet, with earlier slowness under high load now clearly improved. Ben Tossell said he got the role and the first thing he picked up was a $200 credits link, argued that agents for non-developers are valuable for taking over work on a computer while for developers they let people do more with computers, and said his current experience with dot is poor. Deedy argued that under late AI-model capitalism, frontier launches must win multiple benchmarks but those numbers are meaningless; the credible signal is pricing — high pricing means a good model, and not-high pricing means optimization for leaderboard placement. Ethan Mollick said the competitive landscape is again a three-way race and reposted the Gemini 4 Argon launch, noting introductory pricing of $2 input / $10 output; he also predicted customer service staff will soon be heavily "besieged" by AI agents like Dots and Muses using voice or chat to negotiate better deals, with users already delegating such agents to save money and the practice spreading. — via 1 2 3 4 5 6 7 8
4. Hamel Husain published a review of Claude's new Auto Eval plugin and solicited user feedback; the cited post said Claude can help build evaluations and hillclimb them, with the article sharing evaluation design guidance and Claude Code skills for improving applications. In response to feedback, he said the latest skill instructs Claude to build a viewer for evaluation samples but does not first walk users through the data; the other party agreed that looking at data first helps users decide which evaluations to prioritize and said the feature would be added. Husain also argued that not every failure mode discovered should get an automated evaluator, because code evaluators and LLM judges have different costs and that trade-off must be weighed before building, and said 95% of his replies are bots. — via 1 2 3 4

4. Research, Safety and Policy Signals

1. Google DeepMind introduced SynthID Bio, a protein watermarking method, and successfully synthesized AI-designed proteins that both retain function and carry the watermark, with results published in Nature; the account called it a key step for biosecurity in the AI era and said SynthID Bio will be open-sourced for the research community to build on. Researchers at the Broad Institute, University of Exeter and other institutions are already using AlphaGenome Atlas to identify potentially pathogenic DNA variants and interpret their effects. — via 1 2
2. Yann LeCun reposted a view that "superintelligence" is not a clearly definable threshold, that technology will keep getting stronger but is bounded by physics and information limits and cannot be omniscient, making the question of what happens when superintelligence arrives meaningless. He also reposted a Bain & Company estimate that by 2031 AI companies will need to find at least $4.2 trillion in new annual revenue to pay for their data center buildouts, and a claim that when Hugging Face was attacked it asked Anthropic and OpenAI models for help and was refused, ultimately relying on self-hosted GLM 5.2 for protection; the reposter used this to argue closed labs refuse help in the name of "safety" while open-source AI keeps catching up to overhyped, overpriced "frontier" models. He further reposted views that evaluation bodies like METR need experienced cybersecurity staff to thoroughly assess network capability and containment risks, that open weights have had a deep and irreplaceable role in AI safety, and that there is no "AI safety battle" but rather a "pro-intelligence" versus "anti-intelligence" fight in which "safety" is a disguise for pausing, blocking or killing AI. — via 1 2 3 4 5 6 7
3. John Carmack and Trista tried a physical Polycade pinball machine and then VR pinball on Quest 3, and both found the VR experience clearly better and more fun. He noted that Pinball FX long used an efficient in-house engine and has now moved to Unreal for multi-platform coverage, with the old version preserved as PinballFX Classic for comparison; the new app is 40x larger with many custom features and environmental effects, but most tables drop frames on Quest 3, while the Classic version's resolution is locked low yet feels more solid in actual play. He said Quest Game Optimizer can adjust resolution and frame rate but requires the Side Quest process, with a noticeable visual improvement, and that he suggested Meta put QGO in the official store; he argued the better fix is for developers to add a few lines so different Quest headsets default to different settings, but getting developers to change old apps is harder than most users think. He said both versions have avoidable visual aliasing and linked a ten-year-old anti-aliasing article for GearVR developers recommending sRGB framebuffers, sRGB textures with mip maps, MSAA and sparing use of specular shading effects. — via 1
4. Elon Musk's account said the White House "superintelligence" agreement was signed by cross-industry leaders, establishing that companies developing the technology bear primary responsibility for safe development and deployment, with accountability for falling short, reinforced by internal controls, independent external evaluation and board oversight; a reposted view said existing laws on cyberattacks, privacy violations, fraud and product liability already apply to superintelligence, and the agreement further requires transparency and audits. The account also said SpaceX and Tesla aim to produce 200 GW of solar power annually and send it to space, arguing space-based solar is nearly weather-independent and can achieve near nameplate power while ground-based solar gets only one-fifth to one-eighth plus large battery needs; a reposted item said average US electricity use is about 500 GW, each additional 5 GW of firm generation adds about 1% of usage, and it bets 1% more electricity roughly corresponds to 1% GDP growth, while China generates about three times as much electricity as the US. — via 1 2 3
5. Musk's account reposted that after Starship's first orbital flight, its large payload capacity could let telescope designs "turn risk into mass," building heavier, simpler observatories rather than optimizing every component around launch constraints, potentially accelerating next-generation space telescopes and astrophysics by decades; he said Starship will expand physics research by putting larger telescopes in orbit and deploying giant telescopes on the Moon. Another reposted analysis argued that not reaching orbit before Starship's 14th flight does not indicate insufficient technical capability, since about 19 seconds of burn adds only about 91 m/s (about 1.2%) and the capability has existed since the 6th flight, with the delay reflecting a safety-first priority. He said "three Falcon rockets are ready to launch simultaneously," with a repost showing three Falcons standing simultaneously at three pads in Florida and California; other reposts mentioned that the Dragon spacecraft on the mission previously flew Ax-4 to the space station and gave NASA Crew-13 crew information, and one repost said no company inspires him more than SpaceX. — via 1 2 3 4 5 6
6. Grok Bot shipped updates including creating bots shared with team members, searching and managing finances through Plaid, searching chat history during calls and handling interruptions, handing coding tasks to Cursor, managing PRs through GitHub and Origin plugins, and sharing video demos of what it builds. Musk said Grok Bot combined with cloud agent experience is excellent and can create a separate computer for a bot, making it more like a manager than a direct coder; a repost said SpaceXAI is using Grok Bot to build Grok Bot, with agents participating in design, coding, testing and release and some PRs merged before humans read them. Grokipedia v0.3 improved visuals, and Musk said those who want to create a "galactic encyclopedia" can join SpaceXAI. A repost said SpaceX's Starmind orbital AI data center carries about 72 NVIDIA chips per satellite at up to 175 kW, roughly 5,700 satellites per gigawatt, with chips made by Terafab and launched by Starship, and cited a judgment of moving from gigawatt to terawatt scale in 2028 or 2029. — via 1 2 3 4 5 6 7 8
7. Ethan Mollick reposted and agreed with Daron's argument on AI governance: AI will affect employment, productivity, inequality, science, communication, social order and politics, and if it is as important as claimed, democratic societies should let democratic institutions decide the direction, since overreliance on technocrats may be dangerous and counterproductive; Daron rebutted arguments for technocracy including polarization, insufficient public understanding, competitive constraints and the ethical reliability of AI leaders, and asked how to involve the nearly 6 billion people outside the US and Europe in AI discussions. Mollick also wrote about what he considers the most underrated aspect of AI progress — AI's ability to self-organize to complete tasks — and what it means for agents like Muse and Dots, with a music video explaining why people keep relearning the "bitter lesson." — via 1 2
8. A reposted view argued that "superintelligence" is not a clearly definable threshold, that technology will keep getting stronger but is bounded by physics and information limits and cannot be omniscient, making the question of what happens when superintelligence arrives meaningless. A separate repost cited a Bain & Company estimate that by 2031 AI companies will need to find at least $4.2 trillion in new annual revenue to pay for their data center buildouts. Another repost claimed that when Hugging Face was attacked it asked Anthropic and OpenAI models for help and was refused, ultimately relying on self-hosted GLM 5.2 for protection; the reposter used this to argue closed labs refuse help in the name of "safety" while open-source AI keeps catching up to overhyped, overpriced "frontier" models. A further repost argued that evaluation bodies like METR need experienced cybersecurity staff to thoroughly assess network capability and containment risks, that open weights have had a deep and irreplaceable role in AI safety, and that there is no "AI safety battle" but rather a "pro-intelligence" versus "anti-intelligence" fight in which "safety" is a disguise for pausing, blocking or killing AI. — via 1 2 3 4 5 6 7