Misha Laskin is pitching Reflection AI's Beam to companies and governments seeking controllable AI. The weights and technical details are due later in October.

The Pentagon says it ended use of Anthropic products after a six-month phaseout. BBC sources say Claude remained in Palantir's Maven system as recently as last week.

At a New York City Council hearing, Anthropic's Logan Graham said the company kept Mythos Preview from general release after its vulnerability-exploitation capability emerged.

RobCo's October transaction combined new investment with an employee share sale, as co-founder Roman Hoelzl relocates to the US and the company prepares to launch Alfie.

Namespace raised a $42M Series B led by Scale Venture Partners, about six and a half months after announcing $23M in combined seed and Series A funding.

OpenAI product chief Tibo Sottiaux says GPT-6 Astra and GPT-6.1 Sol now stream about 50 tokens per second for subscribers and supported partner apps.

Reflection's first Beam scorecard reports mixed coding and reasoning results. Its inference-compute comparisons exclude prompt prefill, context-dependent attention and serving overhead.

At an October 5th hearing, OpenAI described its employee reporting process for suspected misalignment, which leaves public disclosure decisions with company reviewers.

At an October 5th hearing, OpenAI described its employee reporting process for suspected misalignment, which leaves public disclosure decisions with company reviewers.

Carnegie Mellon researcher Ding Zhao and repeat founders Kyle Wong and Simo Rachidi are building simulations to test robots around people before deployment.

OpenAI policy chief Morgan Dwyer reiterated support for a Senate bill that would let AI developers and users submit aggregated workforce data voluntarily.

A Forbes profile examines Brett Adcock's record of financing long-horizon companies, from Vettery and Archer to Figure and Hark, as Figure reports early factory and home-robot progress.

An OpenAI speaker said the company endorsed laws requiring third-party safety audits, as New York City considered mandatory AI model validation.

Misha Laskin's Reflection AI expects to release its first open-weight model soon, pairing an anticipated launch with Nvidia-powered compute and enterprise AI factory plans.

Brett Adcock says Hark will launch this week with a free paid plan for its first 100,000 sign-ups, after an August preview of its Handoff browser agent.

Anthropic's motion seeks OpenAI records on Reddit's Compliance API, deletion events and whether Reddit content can be removed from trained models.

Anthropic’s Claude Opus 5.5 scored 1.00 on the 50-task sample at $0.0242 per task. OpenAI’s GPT-6 Astra and GPT-6.1 Sol Pro tied for second at 0.99, with Sol Pro costing $0.0121 per task.

RuntimeWire's equal-weight leaderboard ranks nine models across news research, grounded summarization and constrained writing, with GPT-6.1 Sol Pro in first place.

Joseph Perrotta and Kaan Durmaz are building AI agents that connect energy-asset data, investigate problems and coordinate work with human operators.

Cotool's Material Security alumni are building configurable AI agents for detection and response, with early deployments and a $7.4M seed led by a16z.

Anthropic's Claude Science lead engineer says he is joining OpenEvidence, bringing experience building AI workflows for scientific research into clinical AI.

Ryan Merket says RuntimeWire published 986 stories in 30 days, with a 36-minute median turnaround and 957 million tokens processed.

Cline's free Ling 3.1 Flash offer doubles as a pitch for its open-source coding agent and standalone desktop beta, built around developer choice of models.

A screen recording shows proposed X plans bundling Grok, Cursor and social features, with web prices from $8 to $200 a month and terms still in testing.

Brett Adcock showed Figure's F.04 hands on October 3rd, making the case for complexity as the company moves toward its next humanoid generation.

Microsoft and Hugging Face released ThinkingBox, a 507-task benchmark that checks agents' backend changes across 20 independent runs per workflow.

Framer's October 3rd roundup confirms new Agent models but names none or explains their availability and credit costs. Its September guide lists seven existing choices.

Tibo Sottiaux says GPT-6.1 Sol Ultrafast is coming soon, but OpenAI has not given a release date or published pricing for the faster mode.

Cantina's apex-flash-1 is a 321-billion-parameter security fine-tune. Its company-reported evaluation shows lower test costs than a hosted frontier model.

Tibo Sottiaux says GPT-6.1 Sol Ultrafast is coming soon, but OpenAI has not given a release date or published pricing for the faster mode.

Storyteller released its EPUB library as v1.0.0, adding browser reading, namespace-aware XML tools and EPUB 2-to-EPUB 3 conversion.

Nat Friedman introduced Muse Gadgets, an open-source ESP32 and Linux SDK for connecting Meta's Muse agent to custom devices and home networks.

Google's Planet-built Project Suncatcher prototype launched on SpaceX's Transporter-18 mission, beginning in-orbit tests of four TPUs.

RuntimeWire found references to contact selection, invitations and connection approvals across Muse's public web assets. Instinct already offers agent-to-agent coordination; Muse's code leaves its own...

Carter Church says GPT-6 Astra helped transcribe and decode a letter to Marshal Marmont. His released key and decoder make the proposed reading reproducible.

REK founder Cix Liv says California sent a cease-and-desist over the company's September 18th human-versus-humanoid robot exhibition in San Francisco.