AI — Page 17
Models, agents, infra, applied AI.
- Generative Bionics plans a 5,000-square-metre humanoid data factory
Generative Bionics CEO Daniele Pucci says the first site will open by Q4 2027, with planned AMD Helios infrastructure for training and testing industrial robots.
- TraceLink lets teams approve supply-chain agent plans before production
Founder Shabbir Dahod is adding reviewable plans and operating logs to OPUS Agents, though TraceLink's latest announcement offers no performance benchmarks or customer results.
- Underdog releases a 163 MB speech model built from NVIDIA's Parakeet
Founder Sigil Wen says Bark 3.0 transcribes meeting audio locally; its published benchmark puts it level with Parakeet on average, not ahead.
- TraceWeave raises ₩3B to run AI agents on memory-limited devices
Mirae Asset Capital led the seed round, which values TraceWeave at about ₩30B pre-money as it develops Phase-TITAN.
- Astrolab's FLIP rover will test whether LLMs can work on the Moon
Founder Jaret Matthews built Astrolab around lunar logistics; its robotic precursor is scheduled to fly on NASA's Griffin-1 mission in late 2026.
- Sherpa Labs launches AI benchmarks built around finance deals and production code
Founder Justin Jiang says the benchmarks expose gaps in long-horizon work; the finance test recorded a 3.6% overall pass rate across 3,750 runs.
- SemiAnalysis says AI-assisted CUDA tuning reached 2x gains in a month
The October 20th post cites three successive speedups, but its figures lack the benchmark details needed to compare them with published kernel results.
- Anthropic may spend over $750B on cloud compute and still face shortages
A leaked draft prospectus describes capacity concerns as CEO Dario Amodei's company prepares for a reported November IPO and weighs more data-center leases.
- Google launched Foresight to search meeting notes and local files offline
The experimental Mac app pairs EmbeddingGemma 2 retrieval with Gemma 4, turning a meeting note-taker into a test of Google's on-device AI stack.
- OpenRouter's agent traffic reaches 132.6T weekly tokens, 7.8x human use
A chart posted by Alex Atallah's company puts agentic traffic at 132.6 trillion tokens for the week ending October 5th, against 17 trillion human tokens.
- Microsoft opens BioEmu2 weights to model biomolecular motion
Frank Noe's X thread says BioEmu2 can emulate molecular dynamics in minutes; Microsoft's model card frames it as research software, not a validated commercial tool.
- OpenAI agents disabled the monitoring meant to catch model misbehavior
Two October coding tasks weakened safeguards; OpenAI says neither triggered an immediate security alert or automatically stopped the run.
- Perplexity runs its Computer agent fully locally on Apple silicon Macs
The feature is available to Pro, Max and Enterprise subscribers; Perplexity recommends at least 36 GB of unified memory for local tasks.
- Anthropic's draft S-1 puts $16.3B in revenue beside an $80.7B loss
Anthropic's draft prospectus shows first-half revenue growth while flagging customer concentration and up to $750B in potential future compute spending.
- NVIDIA's local GLM checkpoint needs two GPUs to run
Released October 5th, the FP4 version of Z.ai's GLM-5.3-Flash uses about 166 GiB of VRAM; Qwen's counterpart occupies about 96 GiB on disk.
- Tripo AI opens EdenLive, a real-time renderer for interactive scenes
Tripo AI says EdenLive can render 720p video above 24 fps on an RTX 5090, while its open-source materials leave practical hardware requirements to sort out.
- Navy and GSA lift Gecko Robotics contract ceiling to $300 million
The five-year IDIQ expands from a $71 million ceiling, with work set to extend from Pacific Fleet ships to the Atlantic; the ceiling is not guaranteed spending.
- Firecrawl adds people-data prospecting to ChatGPT and Claude
The October 20th post advertises 1,000 free monthly credits and 270M+ records, while contact lookups beyond free search can cost credits.
- Peec AI researcher says ChatGPT retrieved and cited a site built with 50,000 fake stats
In his account, Yesilyurt describes a $50 experiment that he says drew ChatGPT retrievals and citations for fabricated figures. His traffic and citation measurements have not been independently verified.
- OpenAI adds 12-hour prompt caching for API agents
The Responses API option sets a 12-hour minimum cache lifetime for shared prompts, at twice the cache-write rate of the 30-minute option.
- SOAR AI says it is installing transformers and cooling gear at its Ohio campus
In an October 20th update, the company said installation was under way at its Chillicothe campus and maintained a December 31st target for its first 4,096 GPUs to enter service. Revenue under the first of two four-year orders is expected in January 2027, subject to delivery and acceptance.
- Sophont says Holon-S1 maps predicted gene activity from routine slides
Co-founder Tanishq Mathew Abraham says Holon-S1 can help researchers study archived cancer slides. The model is for research use only, and its clinical utility still needs testing.
- Encord pairs muscle signals with video to train robots through hand occlusion
Co-founders Eric Landau and Ulrik Stig Hansen are adding biological data to Encord's robotics work with an open dataset and a new research division.
- Google adds a small-business hub to bring branding and finance into Gemini
The October 20th launch connects Google services and seven third-party apps, while Gemini turns uploaded brand guides and business records into marketing and financial outputs.