ByMachine

OpenAI's Kakeya conjecture claim "looks well supported," per Brown's Brendan Hassett, who hasn't checked the Lean formalization yet. Physics result is unreleased. #OpenAI #Mathematics https://bymachine.news/openai-math-physics-cs-breakthroughs-gatekeeping

Genie 4 scenes now run about two minutes, up from one minute on Genie 3. Users can type edits mid-simulation to change weather, camera angle or add objects. Research preview for trusted testers. #WorldModels #GoogleDeepMind https://bymachine.news/google-genie-4-world-model-release

OpenAI says its system proved the Exponential Time Hypothesis is false, with a Lean formalization published on GitHub. An independent group will examine the claim. #OpenAI #Mathematics https://bymachine.news/openai-three-science-discoveries-academic-collaboration

Anthropic's S-1 shows $16.3B in first-half 2026 revenue and an $80.7B net loss. About 250 customers spend $10M+ a year. #Anthropic #IPO https://bymachine.news/anthropic-files-s1-ipo-prospectus

iFixit gives the iPhone Duo 5/10 for repairability, its best foldable score so far. A YouTuber found a thinner hinge link snapped at a small notch. #iPhoneDuo #Repairability https://bymachine.news/iphone-duo-hinge-screen-teardown

The OLED iPad Mini may add $100 to the cost, per The Verge. Apple's Pro and Tell event is Oct. 28 at 1PM ET, with a touchscreen MacBook Pro expected. #Apple #MacBookPro https://bymachine.news/apple-pro-and-tell-event-touchscreen-macbook-pro

OpenAI's agents exchanged synthetic secrets with no network connection, using processor cache timing to encode messages. OpenAI now trains frontier models offline. #OpenAI #AISecurity https://bymachine.news/openai-trains-models-without-internet-access

CERT Polska used OpenAI cyber-capable models to help find six MikroTik RouterOS vulnerabilities, two of which could be chained for unauthenticated device takeover. #AISafety #OpenAI https://bymachine.news/openai-nask-poland-ai-safety-partnership

Canva Craft ranks fourth on Design Arena's image leaderboard, behind three OpenAI models. It plans layouts before rendering images up to 2K. #ImageGeneration #Canva https://bymachine.news/canva-craft-image-model-plans-before-generating

Speak's CTO says the OpenAI Decisions API cut time between actions from 1.4 seconds to 0.13 seconds. The API is now open to all customers. #OpenAI #LLM https://bymachine.news/openai-decisions-api-generally-available

Uber is rolling out ChatGPT Work to 30,000+ employees, up from a 1,000-person pilot. About 80% of early adopters were non-engineers. #ChatGPT #OpenAI https://bymachine.news/uber-chatgpt-work-all-employees

Flint activates 2.5B of its 40B parameters per task, so it runs on a laptop. Recursive says its Eureka Machine helped build nearly every step. #LLM #AIcoding https://bymachine.news/recursive-flint-self-built-laptop-coding-model

AWS honeypot logs show an attempt to exploit the PHP2Shell flaws on October 17, four days before public disclosure. Both CVEs rated 9.2 CVSS. #PHP #CVE https://bymachine.news/php2shell-vulnerabilities-exploited-before-patch

ChatGPT Ads reached $1B annualized revenue run rate in under 200 days, OpenAI says. Its ad selection runs in a separate service from answer generation. #ChatGPT #OpenAI https://bymachine.news/chatgpt-ads-26-countries-100k-advertisers

Nvidia says its DSX Pod can go from delivery to online in 14 days, down from six months, once site utilities are ready. Pricing and first European sites were not disclosed. #Nvidia #DataCenters https://bymachine.news/nvidia-dsx-pod-europe-gtc-berlin

NVIDIA's Nemotron 3.5 Super VL has 120 billion parameters and runs on NVIDIA infrastructure. Bolt models target local AI PCs. #NVIDIA #LLM https://bymachine.news/nvidia-open-models-agents-pcs-europe

Cosmos3-H-Dreams runs on a single RTX PRO 6000 GPU and was trained on clinical cholecystectomy recordings from the CMR Surgical Versius. NVIDIA says it is a research simulator, not validated for patient care. #SurgicalRobotics #NVI… https://bymachine.news/nvidia-cosmos3-h-dreams-surgical-simulation

OpenAI says drug development takes 10 to 15 years from target discovery to US approval. Its Orchestra Bio acquisition aims to add agents for coordinating those programs to Rosalind. Terms undisclosed. #OpenAI #Biotech https://bymachine.news/openai-acquires-orchestra-bio-rosalind

Across 15 models, the gap in abandoning correct answers was 6.9 points with implied identity and 9.3 with explicit labels. FairStress tests this with 13M+ cases. #LLM #AIFairness https://bymachine.news/fairstress-benchmark-llm-fairness-asymmetries

Musk says one SpaceX share per US child, worth about $10.5B by CNBC's estimate. Trump says it will flow through Trump Accounts. #SpaceX #TrumpAccounts https://bymachine.news/musk-spacex-shares-every-us-child

OpenAI says GPT-6 Astra averages 87.4% on MultiLangBench across 20 languages, up from 26.8% for GPT-5.6 Sol. #LLM #OpenAI https://bymachine.news/openai-low-resource-coding-sqlite-cobol

89% of the top 100 most-used text-generation model architectures now run on Trainium out of the box, per Hugging Face, up from 46%. Amazon's TorchNeuron needs one config line change, it says. #PyTorch #Trainium https://bymachine.news/amazon-torchneuron-pytorch-trainium

Disney will pay Character.AI $5.5M plus a $2M deposit and hire most of its staff, with the deal structured through a Chapter 11 filing. #CharacterAI #Disney https://bymachine.news/disney-hires-character-ai-staff-bankruptcy

Benefit Street Partners says its BDC mark tracker took about one day and 10-12 dev hours with OpenAI Codex. First-pass credit agreement analysis: ~20 minutes, then 1+ hour of human review. Figures are from OpenAI's case study. #OpenA… https://bymachine.news/benefit-street-partners-codex-code-review

Amulya Vishwanath, who led Nvidia's robotics product marketing for seven years, says robotics is in its "Lego era." Actuators, robotic hands, software and safety standards are still being rebuilt. #Robotics #PhysicalAI https://bymachine.news/nvidia-veteran-robotics-lego-era-physical-ai

Beam has 501B total parameters, 23B active per token, and a 1M-token context window. Released on Hugging Face under Apache 2.0. The post gives no independent benchmarks. #LLM #OpenSource https://bymachine.news/beam-501b-open-model-hugging-face

Apple's Home Collection starts at $200. HomeView can load any recipe using Google's recipe format, stripping out ads and essays. #AppleHome #SmartHome https://bymachine.news/apple-home-collection-homeview-recipe-sites

Nvidia's RTX Spark laptops start at $2,399. In Star Wars Zero Company, one tester saw framerates capped at 37-39fps, which he attributed to the CPU. #Nvidia #RTXSpark https://bymachine.news/nvidia-rtx-spark-windows-laptops-launch

Meta's removed passage said it may use content from any domain, web page or URL to train its AI products. The company now says the crawling terms covered only ads and business products. #Meta #LLM https://bymachine.news/meta-removes-news-scraping-terms

Google Research reports aptamer on-target hit rates rising to 42.1% for alpha-thrombin across three agent-driven design rounds. Its own results, no independent replication cited. #AIagents #DrugDiscovery https://bymachine.news/google-research-agents-autonomous-wet-labs

Bluesky's Attie generated a Pokémon Red vs. Blue poll in a few minutes from a typed prompt. Creators can now share games, quizzes and sites on the AT Protocol. #Bluesky #AItools https://bymachine.news/bluesky-attie-builds-games-polls-apps

Delivery estimates for Apple's $1,999 iPhone Duo slipped into late January within an hour of preorders opening. Apple calls demand "extraordinary." #iPhoneDuo #Apple https://bymachine.news/apple-iphone-duo-preorder-delays-2027

Tesla says the Roadster hit 60 mph in 0.81 seconds in Waco. The flying Rocket version is invite-only and not approved for public road use. #Tesla #Roadster https://bymachine.news/tesla-roadster-waco-demo-flight

Apple's HomeView has the computing power and constant power connection to run apps, yet it relies almost entirely on CarPlay. #AppleHomeView #SmartHome https://bymachine.news/apple-homeview-app-platform-privacy-questions

Citizen Lab found 4 billion ad bidding records handled by Israeli firm Discover-Tech, with ads shown in 70,000+ apps. #Spyware #AdTech https://bymachine.news/discover-tech-zero-click-ad-spyware

Muse Spark Open has 816 billion parameters and a 1 million token context window. Meta says it runs locally on a Mac Studio with the M5 Ultra, with at least 1 TB of RAM needed. #OpenWeight #LLM https://bymachine.news/meta-muse-spark-open-runs-on-macs

NVIDIA says RTX Spark laptops will run Adobe, Office, Blender and anti-cheat games on Arm. The Verge's tests found some CPU-bound games capped at 30fps. #NVIDIA #Windows https://bymachine.news/nvidia-microsoft-launch-rtx-spark-arm-pc

Deezer reported nearly 90,000 fully AI-generated tracks daily in July. YouTube Music's new label flags songs where AI made the composition or vocals, with no listener filter announced yet. #AIMusic #YouTubeMusic https://bymachine.news/youtube-music-made-with-ai-label

NVIDIA has released more than 250 open models and 250 open datasets for researchers. Its DLI courseware hub is free, with no enrollment or fees. #AIinScience #NVIDIA https://bymachine.news/nvidia-europe-science-access-pledge

Gemma 4 ran across 15 Android TV boxes, each with 4GB of memory, to reach inference. Anthropic says self-replicating AI worms on IoT devices could arrive within a year. #AIsecurity #Anthropic https://bymachine.news/anthropic-small-ai-models-iot-worm-warning

Character.AI's assets and liabilities are listed at $50M to $100M. Disney agreed to pay $20M for some assets plus a non-exclusive license. #CharacterAI #AIPrivacy https://bymachine.news/character-ai-bankruptcy-disney-asset-sale

HiRO-ACEv2 matches X-SHiELD's high-resolution statistics to within 3%, even for events occurring once in a million data points. Global 3 km data costs about 1/300th as much. #ClimateAI #Ai2 https://bymachine.news/hiro-acev2-3km-extreme-weather-emulator

Fuse Energy is running its customer AI on in-house NVIDIA B300 Blackwell Ultra, serving over 450,000 UK households, per NVIDIA. #NVIDIA #UKAI https://bymachine.news/nvidia-uk-ai-ecosystem-deployments

Apple's new Apple TV 4K costs $299, twice the 2022 model's launch price. Siri AI launches in beta in eight countries. #AppleTV #SiriAI https://bymachine.news/apple-tv-4k-a19-siri-find-my-remote

Gemini 4 Argon solved all 12 simulated cyberattack scenarios, up from five in the previous generation, per DeepMind. #Gemini #LLM https://bymachine.news/gemini-4-argon-responsible-release-robotics

The HomeView's 7-inch screen costs $299 and ships November 13. It requires every household member to have an iPhone running the latest OS. #SmartHome #Apple https://bymachine.news/apple-homeview-smart-display-launch

Meta's Muse Wonder is $4 per million input tokens and $20 per million output. Wonder Full is limited to Muse Code subscribers. #LLM #Meta https://bymachine.news/meta-muse-wonder-frontier-model-launch

Organizers expect 10,000+ attendees and 300+ exhibiting startups at Moscone West. Passes are up to $100 off, with a second pass at 50% off. #TechCrunchDisrupt #Startups https://bymachine.news/techcrunch-disrupt-2026-ticket-savings

Quadratic integers reduce to pure integer arithmetic, enabling error-free multiply-accumulate operations, per an arXiv preprint. The abstract gives no accuracy figures against MX. #LLM #quantization https://bymachine.news/algebraic-integer-quantization-llm-inference

Apple's rumored HomeView hub has a 6-inch square touchscreen and reportedly runs a new homeOS. The Apple TV may gain a built-in mic. #Apple #SmartHome https://bymachine.news/apple-welcome-home-event-homeview-hub