AI — Page 4
Models, agents, infra, applied AI.
- Google adds two TPU variants as Vahdat takes the Nvidia fight wider
Bloomberg reports two new Google TPU variations aimed at confidential computing and broader information access for models, as Vahdat pushes TPUs beyond Google Cloud.
- Meta launches Muse Ember and prices it at $0.10 per million input tokens for agent volume work
Artificial Analysis data backs Meta's pitch: at $0.10 per million input tokens, Ember is presented as cheaper and smarter than Haiku and Luna on cost per task, with Wonder planning and Ember handling the volume.
- Anthropic reportedly agreed to $3.4B gas generator deal in September
Dario Amodei's Claude lab reportedly agreed in late September 2026 to take over OnePWR's Caterpillar order for up to 1.7 gigawatts, with the exact agreement date still unverified.
- Google puts Maps Platform inside AI Studio with one-click setup
Abinaya Selvam's Maps team moved API keys and billing setup inside AI Studio, with a free sandbox capped at 100 calls per API per day.
- Netflix proposes a synopsis fact-checking cascade and reports 52.4% lower show-level cost
Netflix TechBlog authors Shreya S Ramanujam, Yongce Li, Afshine Amidi, Sina Ghiassian and Benjamin Klein describe a proposed synopsis factuality cascade, and Netflix reports a 52.4% show-level cost reduction in its evaluation.
- Edgevana says it signed over $700M in Blackwell infrastructure deals
Ryan Fay and Subhan Jahromi say Edgevana has signed multi-year take-or-pay deals for NVIDIA B300 and GB300 systems, alongside $7 billion of contracted demand.
- Plato CEO says his OpenAI agent sent 300 emails about an office move
Plato CEO Robert Farlow handed his office move to an always-on agent, then shut it down and went back to reviewing drafts himself.
- Israel wants 300,000 GPUs and its own frontier model to run its AI stack
Erez Askal split the target into 100,000 GPUs for domestic use and 200,000 for companies and data centers, and tied both to two gigawatts of power.
- AI data center builders turn to used turbines to bypass long waits
Reuters says new U.S. turbines take six years on average, pushing Crusoe and others to ProEnergy conversions, FTAI Power and Boom's unproven Superpower.
- Flashduck lets an AI improvise its poses instead of playing canned animations
Shanmao Tech founder Xiao Qiaowei built Flashduck on his Bot 1 kit, with DuckClaw giving a model direct control of six servos through DingTalk.
- Venice adds Krea 3 with eight-image edits and anonymous access from $0.04
Venice lists Krea's proprietary editor with 1K to 4K output from $0.04, anonymous access, and up to eight reference images per generation.
- Legora adds 250 prebuilt Skills to its legal AI agent
CEO Max Junestrand says the tools encode firm-specific review, negotiation and communication preferences in plain language, with access for all customers.
- Nous Research merges saved-login access for Hermes Desktop's browser
Nous Research's November 7th merge lets Hermes Desktop reuse signed-in browser sessions or fill logins from Hermes Vault, Bitwarden and 1Password.
- Elvis and Eminem rights holders sue Suno over AI voice cloning
Filed November 6th, 2026, the Presley and Eight Mile Style cases challenge Suno's safeguards against artist imitation.
- Diamond Foundry grows a 200mm single-crystal diamond aimed at AI chip heat
Diamond Foundry CEO Martin Roscheisen announced the unpolished crystal on November 5th, after scientists Stefan Gsell and Martin Fischer spent 25 years working on diamond growth.
- DeepSeek plans plugin marketplace as Harness reports 1.54 million daily users
Liang Wenfeng's DeepSeek disclosed the figures on November 5th. The Harness count covered users of the official API, while plans for a marketplace and author support have no announced launch date.
- OpenMed's synthetic discharge demo rejects a false potassium claim
Maziyar Panahi's OpenMed used a planted potassium error in a synthetic note to show a brief returning no text rather than passing a contradiction to review.
- Asahi Linux plans GPL switch after alleging AI-assisted code copying
The volunteer project founded by Hector Martin said on November 7th that it has begun moving its repositories to GPL version 2 or later after alleging that code was copied from personal forks, sometimes without attribution or rewritten with LLM assistance.
- WIRED co-founder Kevin Kelly agrees: Its AI coverage has lost its internet-era optimism
Ethan Mollick says his Astra tests found differences between WIRED's AI and early internet coverage. Co-founder Kevin Kelly replied: "That feels about right."
- Nous Research ships Hermes v0.23 with one gateway for every conversation
Released November 5th, Hermes v0.23 puts CLI, Desktop, editors and bots on one gateway and labels uncertain claims in compaction summaries as UNVERIFIED.
- Tencent says DSpine speeds Qwen3.8-27B inference in agent tests
The five-layer speculative-decoding drafter delivered 2.71x to 3.84x the autoregressive baseline in Tencent's agent throughput tests.
- Let's Encrypt AI paper sets 2027 goal for confidential AI chats
An independent position paper with no named author or publication date calls for confidential computing to protect AI chats on most AI GPUs by the end of 2027.
- Fireworks says Nexus costs four times less than a single frontier model
Fireworks AI co-founder Lin Qiao's Nexus routes routine turns to open models and reserves closed models for harder work. The 1,222-task claim is company-reported, and the X post supplied for this story carries no verified publication date.
- CoreWeave seeks Texas data center cooperation from governor-elect Hinojosa
Michael Intrator says CoreWeave supports growth that works for Texas; Gina Hinojosa campaigned for a pause on new data centers.