AI — Page 28
Models, agents, infra, applied AI.
- National Compute launches shared AI grid with 760 MW connected or in sight
Anjney Midha's National Compute will pool capacity across providers and chip types, aiming to widen access for startups, universities and public-sector researchers.
- TII says its 1.6B Falcon-ASR beats larger models on its Emirati test
The Abu Dhabi research institute reports a 22.73% word error rate on its internal Emirati test; the model also supports four languages beyond Arabic.
- Contamos lets AI assistants write to a shared, multi-currency ledger
The Uruguay-operated product supports Claude and ChatGPT, but its records depend on users entering transactions or reviewing imported bank statements.
- WEKA Federal clears a faster procurement path for NeuralMesh
The designation puts WEKA's AI data platform in a Department of War marketplace, where buyers can consider it without repeating the initial technical evaluation.
- SpaceX says Grok Bot will use rival AI models for different tasks
Elon Musk named Claude Opus 5.5, Midjourney and Suno, but gave no details on how Grok Bot will choose providers or handle those services' costs and data flows.
- Claude Opus 5.5 user says an AI agent deleted his C: drive
Zac Cook says daily backups to a Synology NAS saved his data; his October 6th post does not establish which tool or command caused the deletion.
- Anthropic is building a writing assistant that puts Claude inside other Windows apps
RuntimeWire found a hidden “Quick Write” interface for rewriting selected text and drafting at the cursor. The implementation ships in Claude’s Windows package; its rollout remains unconfirmed.
- Oki Home puts local AI and a searchable digital life in a $1,799 PC
Luofei Chen's startup is taking $59 reservations for a compact computer that stores personal files locally and is scheduled to ship in the U.S. in December 2026.
- micro1 tests whether four personal agents can be trusted with real accounts
Ali Ansari's benchmark scores task success separately from permission, privacy and accuracy failures. Its early leaderboard is based on 40 scored attempts.
- TII trains a 7B Falcon model to answer in Emirati Arabic
The Technology Innovation Institute reports an 84.83% score on its Emirati-language benchmark, which tests dialect, cultural knowledge and poetry.
- Researchers trace Amap-querying agents' code to Tencent Cloud
Swarmchasers found parallel agents probing Amap for venue-entrance data, with infrastructure evidence pointing to Tencent Cloud but not proving who deployed them.
- Leviathan indexes million-row datasets for AI agents without filling context windows
Joshua Gunn's open-source Rust tool returned relevant records in 99% of a synthetic million-row test, using far fewer tokens than the tested grep approach.
- Google launches Nano Banana 2.1 ahead of the older image API's earliest shutdown
The new image model supports 4K output, Google Search grounding and multi-image editing. Google's deprecation schedule lists October 29th as the earliest shutdown date for Gemini 3.1 Flash Image, its recommended replacement.
- Musubi puts its moderation policies inside a 1.7B-parameter model
Co-founder Filip Jankovic says the open-weights release builds on an idea he was exploring before decision models became a talking point; Musubi's own tests show both speed and limits.
- Authors First raises $10M+ to give novelists final cut on AI-assisted TV
Founder Robert Hamwee is betting lower production costs can bring ambitious books to screen, with Conn Iggulden's six-episode Genghis adaptation as the first test.
- OpenAI publishes 722 math manuscripts generated by an unreleased model
OpenAI's collection includes Lean proof artifacts and reasoning summaries. OpenAI says verification varies across the papers and that citations and exposition need further work.
- Vercel CEO Guillermo Rauch wants human-written READMEs and agent-ready docs
His proposal separates public explanations for developers from denser instructions for coding agents, with human editing as the quality bar.
- Anthropic puts Claude inside Google Workspace, with edits users can approve
The beta adds a Claude sidebar to Docs, Sheets and Slides, plus connectors for creating and editing Google files from inside Claude.
- Google releases a 740M-parameter embedding model for local multimodal search
EmbeddingGemma 2 maps text, code, images, video and audio into one vector space; Google says its weights are available under Apache 2.0.
- Myrtle.ai's VOLLO hits 50M tree inferences per second in STAC test
On an AMD Alveo V80LL and Blackcore server, VOLLO delivered sub-2-microsecond p99 latency across three tree models; STAC lists it as ML-20260925, while myrtle.ai's release cites MRTL2026905.
- Grok Bot added Team Bot managers and voice calls in September releases
An X post labels the changes version 0.68.1, while SpaceXAI's changelog lists the features in versions 0.64.0 and 0.65.0.
- Josh Elman says consumer AI agents need a better first step
The a16z partner points to an experience problem; OpenClaw shows how setup, model access and permissions shape an agent's first impression.
- Z.ai adds GLM-5.3 to Amazon Bedrock for enterprise coding agents
AWS launched the model on October 5th with managed APIs, prompt caching and cross-Region inference; access is limited to eligible customers.
- Applied Intuition moves its Nissan work from simulation into vehicle software
The October 5th agreement expands a relationship dating to 2020, putting Applied Intuition's Vehicle OS and development tools alongside Nissan's engineering teams.