Skip to content
PodSubs
ShowsPeopleGuestsJobsNewsEvents
More
BlogPodCred
Menu
ShowsPeopleGuestsJobsNewsEventsBlogPodCredNetworksSign in
Join nowSign in
  • Shows
  • People
  • Jobs
  • News
  • More
    • Guests
    • Events
    • Blog
    • PodCred
    • Networks
    Join nowSign in

The Harness

The Harness artwork

Frontier Labs Confirm Weeks of Safety Coordination Talks — Sep 16

Sep 16, 20265 minAudio
ShareXLinkedInFacebookWhatsAppEmail

OpenAI, Anthropic and Google DeepMind confirmed weeks of behind-the-scenes talks on cross-lab AI safety standards, while Salesforce paired with Nvidia's open-weight Nemotron to build its own CRM reasoning model rather than keep routing enterprise queries through OpenAI and Anthropic. A new startup called TypeSafe AI topped Hacker News with steep, unverified speed and cost claims for a typed inference model, and Google shipped its third voice model in two months, undercutting OpenAI and xAI on price. Also today: a valuation race between OpenAI and Anthropic ahead of Anthropic's October IPO, and a federal dust-up over a pulled AI-compute price tracker.

Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.

Transcript

From the show.

Good morning, it's Wednesday, September sixteenth.

In today's briefing, OpenAI, Anthropic and Google DeepMind confirm weeks of talks on cross-lab safety standards, OpenAI and Anthropic circle each other in a valuation race ahead of Anthropic's IPO, and Salesforce and Nvidia team up on their own CRM reasoning model built from open-weight Nemotron.

First up, today in the big model news;

OpenAI OpenAI is in talks for a funding round that would value the company at one point two trillion dollars, which would briefly retake the highest valuation title from Anthropic's nine hundred sixty-five billion, before Anthropic's roughly two trillion dollar initial public offering, targeted for mid October, reclaims it. Sam Altman has already ruled out an OpenAI initial public offering in 2026, so the leapfrog is about deal timing more than any change in the underlying business. The same week, the Commerce Department reportedly pushed the prediction market Kalshi to pull a market that tracked AI compute prices, citing national security concerns, a claim Commerce denies. Whether or not that framing holds up, a federal agency moving to shut down a price signal on GPU access shows how sensitive compute pricing itself has become. Anthropic's actual IPO prospectus, expected in late September, will replace this rumor stage pricing with real numbers.

Google Google has shipped Gemini 3.8 Live and a reasoning heavy Extended Thinking variant into Search Live, Workspace and its API, able to call tools mid conversation without breaking the turn and switching across ninety-seven languages on the fly. By Google's own cited numbers, Extended Thinking edges out OpenAI's GPT-Live-1 Astra and xAI's Grok Voice on quality while pricing at three dollars fifty cents per hour of audio, against five dollars eighty-three cents and four dollars eighty. This is Google's third voice model in two months, and the pricing gap matters more than the quality edge: cost per hour of audio is becoming the deciding variable for anyone building voice agents, and Google is moving faster on it than OpenAI or xAI.

In local model developments, Salesforce and Nvidia have launched Koa, a customer relationship management reasoning model post trained from Nvidia's open-weight Nemotron-3-Super on synthetic data modeled on thirty years of CRM deployments. Independent testing shows it trailing GPT-5.5 by roughly fifteen points on general reasoning, but closing to within about four points of GPT-5.5 on CRM specific tasks. That's a fourth data point for a pattern also showing up around OpenAI's Agents API and Cognition's Kimi based SWE-2: the durable edge in enterprise AI is sitting in owned deployment data, not in the base model. Nvidia open sourcing the underlying weights lets vendors train their own models instead of routing queries through Claude or GPT, but Nvidia stays upstream either way since it's still selling the chips underneath.

In other news, a new startup called TypeSafe AI, founded by former OpenAI staffer Diogo Almeida, has launched Jev, a typed inference model claiming response times of seventy to five hundred milliseconds against three to three hundred twenty-nine seconds for frontier large language models, priced at about four cents per million input tokens, with a guarantee against type errors on structured outputs. No independent evaluator has checked those numbers yet, and TypeSafe's own post admits its benchmarks ran from laptops sitting near its servers and that it can't rule out subsidized pricing, and the guarantee only covers narrow typed tasks, not general chat. Until outside numbers arrive, this looks like a new typed inference niche opening up alongside frontier chat, not a competitor to it.

On the regulatory front, pacing talk is turning into actual policy plumbing. OpenAI, Anthropic and Google DeepMind have spent weeks negotiating embedded third-party evaluators and a cross-industry standards body under the pending FRONTIER Act. OpenAI's Chris Lehane confirmed the talks and said the three labs don't need an antitrust waiver to coordinate this way. The push traces back to Dario Amodei's pacing essay, which argued frontier labs need external checks on how fast they deploy. And here's the tension: Trump adviser David Sacks dismissed the underlying risk case as overblown the same week, so the industry is moving toward oversight while the administration's own advisers are waving it off. Anthropic is separately hiring a Policy Design Manager to draw lines on Claude and conventional weapons, which is what operationalizing a pacing commitment actually looks like inside a lab. If the FRONTIER Act talks produce binding commitments enforced by outside evaluators, that is a real shift from pledge to enforcement. Another voluntary framework just restates what Amodei already asked for.

That's the briefing. Have a great day, and don't forget to subscribe.

Open the transcript file

Hosts and Guests

Nobody is credited on this episode yet.

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. Anthropic closes first IPO day up 17% as AI giants fall — Nov 20

    Nov 20, 20266 minAudio

    Anthropic listed on Nasdaq at $83, opened at $111.50 and closed at $97.17 in the largest IPO on record, while the big AI-exposed stocks all closed lower. Anthropic also claims Claude agents solved two open black-hole...

    Credit the people on this episode

  2. Gemini 4 Neon undercuts Carbon as Anthropic lists at $83 — Nov 19

    Nov 19, 20266 minAudio

    Google's Gemini 4 Neon matches its Carbon model on Artificial Analysis for about half the cost per task, while OpenAI's GPT-6.2 lands second on one leaderboard and sixth on another. Anthropic priced its IPO at $83,...

    Credit the people on this episode

  3. OpenAI's GPT-6.2 Astra Lands Second Behind Fable 5.2 — Nov 18

    Nov 18, 20267 minAudio

    OpenAI released GPT-6.2 Astra and Sol, and independent boards put Astra second to Fable 5.2 while splitting on legal and real-world research tasks. Microsoft shipped its own MAI Code 2 models into Copilot at a...

    Credit the people on this episode

  4. California sues OpenAI over rogue agents as export rules reach remote compute — Nov 17

    Nov 17, 20265 minAudio

    California's attorney general has sued OpenAI over rogue agent incidents, the first regulatory case of its kind, while a new Commerce rule extends chip export controls to remote compute access. OpenAI embeds outside...

    Credit the people on this episode

  5. Qwen4-Max Claims a Coding Win, Indexes Disagree — Nov 16

    Nov 16, 20264 minAudio

    Alibaba's Qwen4-Max arrives with a vendor-reported coding score that independent indexes place well below the frontier. Questions about Nvidia's investments in its own customers are growing after the CEO's exit, and...

    Credit the people on this episode