AI — Page 15
Models, agents, infra, applied AI.
- Moonshot AI makes Kimi Code ask before trusting project folders
Moonshot AI's October 21st update prompts users to trust newly opened folders, confirms model switches and adds a way to reload sessions after configuration changes.
- Phylo puts Biomni Lab into Bayer's drug-discovery evidence workflow
Phylo CEO Kexin Huang's Stanford research project has grown into an enterprise platform Bayer will use to assemble literature and omics evidence for drug discovery.
- Ai2 releases satellite embeddings that let mapping teams skip repeat model runs
OlmoEarth v1.3 supplies precomputed 10-meter features and a browser-based mapping tool. Coverage is global for 2024 and 2025, near-global for 2023, and limited to selected countries for 2017-2022; Ai2 says its embeddings led two named competitors on most of its evaluations.
- Cloudflare adds app integrations and deployment partners to its AI workspace
Cloudflare's October 19th announcement names integrations including Asana and GitHub, plus three deployment partners. Cloudflare reports substantial internal use; customer adoption and pricing remain undisclosed.
- Higgsfield's AI video guide tells teams to count finished clips, not credits
Alex Mashrabov's company compares six platforms for in-house teams, urging buyers to measure the cost and labor per approved video instead of comparing vendor-specific credits.
- Berget AI opens its emissions method after correcting its Texas estimate
Berget AI revised its Kimi K3 estimate for Texas from 3.9 to 3.2 grams of CO2e per request after reviewing grid data; its method still excludes training.
- Supermicro adds a 256-CPU rack for the work AI GPUs leave behind
Founder Charles Liang is extending Supermicro's rack-scale strategy with a liquid-cooled system designed for agent sandboxes, reinforcement learning and data processing.
- Meta says Muse is launching in Japan
Alexandr Wang said the agent was available to download on October 23rd, weeks after Meta's Japanese launch page said local availability was undecided.
- Microsoft trains Qwen3.8-27B on 187 tasks to improve agent repeatability
The Microsoft and Hugging Face experiment raised average success by 9.08 percentage points across 507 benchmark tasks, while the number of tasks the model solved in at least one attempt fell by seven.
- Canva launches an image model that plans the design before generation
CEO Melanie Perkins started Canva to make design easier for non-specialists. Craft applies that premise to image generation, trained on what Canva says are millions of human-made templates.
- X Chat adds access to xAI's task-running Grok Bot
An October 22nd post from @blankspeaker said the rollout was underway. X's Help Center explains how users connect a bot, assign tasks and review proposed actions.
- Rust off Rails calls 60 Markdown files a web framework and lets AI write the code
The open-source project turns framework conventions into instructions for coding agents, which generate application code owned by the developer.
- Anthropic brings Claude's full platform to Google Cloud's billing and IAM
The Anthropic-operated service adds Claude agents and same-day platform features while letting customers use Google Cloud credentials and commitments.
- OpenAI opens Decisions API in a market TypeSafe's Jev helped create
OpenAI adapted GPT-6 Luna for bounded decisions, entering a category TypeSafe CEO Diogo Almeida says could make software more automatic.
- Podium says its AI agents reached $150M in AI-specific ARR in 2.5 years
CEO Eric Rea says Podium rebuilt its local-business software around agents. Podium reports 20,000 deployed agents but does not define its AI-specific ARR calculation.
- Ideogram launches its first vector model for editable SVG design
Vector 1.0 generates SVGs from prompts or converts images, with API pricing starting at $0.10 per image-to-vector conversion.
- AAIF adds four agent projects, taking its first-year roster to 10
Co-founder Manik Surtani says the Linux Foundation-hosted group now oversees 10 projects, though at least one public onboarding record still showed governance steps underway.
- CaptivateIQ adds AI agents to the math behind quotas and commissions
The agents can propose territory, quota and compensation-plan changes for administrator review. CaptivateIQ says beta customers updated plans 80% faster and saved 16 hours of manual work.
- Higgsfield's AI video workflow reaches German rap videos and TV ads
Producer Oliver Alge used Cinema Studio on projects for German rap artists, a TV commercial and Riyadh Air. The reported results come from Higgsfield's interview with him.
- SOAR AI puts $1.17B in GPU orders against a year-end Ohio deadline
CEO Chris Ensey's new investor presentation maps the buildout and financing plan; the first customer revenue still depends on service starting as planned.
- 1Password lets AI agents sign in without seeing passwords
The Verge reports that users approve sign-in requests on their own devices while 1Password fills credentials. In November 2025, 1Password reported more than $400 million in annual recurring revenue and 180,000 business customers.
- Buildots gives Dot construction tasks to analyze project data
CEO Roy Danon says Dot can investigate forecast delays and verify trade claims against project data. The update is available globally to Buildots customers at no extra cost.
- Hystax open-sources a coding-agent compressor that cut one test bill by 22.7%
CEO and co-founder Nick Smirnov says the tool makes repeated build logs retrievable; Hystax measured savings in a benchmark using Codex and GPT-5.5.
- Hugging Face releases Carbon-X, a DNA chat model without wet-lab validation
Hugging Face released Carbon-X with collaborators in China and Italy; its model card reports benchmark scores, while WIRED says wet-lab validation has not been done.