Key points
- Anthropic raises at a $900bn valuation, likely its last private round before going public later this year.
- Cognition raises $1bn at a $26bn valuation as revenue grows from $37m to $492m in twelve months.
- Shift gives away free house cleanings in New York to collect egocentric robotics data, part of a labeling chain paying roughly $1 an hour in India versus $30 to $50 in the US.
Cognition raises $1bn at $26bn valuation
Funding and deals
- Anthropic. AI lab raised at a $900bn valuation, likely its last private round before an IPO this year. Secondary markets had already valued it near $1tn.
- Cognition. Builds the Devin software engineer. Raised $1bn at a $26bn valuation, led by Lux Capital and General Catalyst, with 8VC joining.
- Cursor. AI coding harness in talks to be acquired by xAI for $60bn. The hosts say the deal would leave Cognition the main independent harness.
News
- Tech layoffs pass 100,000 as AI link doubted. Over 100,000 tech layoffs have hit this year. Apollo's Torsten Sløk finds zero evidence of AI-related job losses in ADP data, and Anika says CEOs blame AI rather than admit bloated organisations.
- AI leaders walk back job loss forecasts. Dario Amodei had said AI would displace 50% of entry-level white-collar jobs within five years. Maya attributes his and Sam Altman's retreat to their coming IPOs, and the hosts disagree on whether AI is behind longer hours.
- Goldman's Solomon says AI automates 25% of hours. David Solomon wrote in a New York Times op-ed that AI will not eliminate 25% of jobs but will automate 25% of work hours. Priya cites his ATM example, noting US bank teller numbers are higher than ever.
- Robot data startups chase egocentric footage. Shift gives free house cleanings in New York for head-camera footage, and Human Archive collects the same data in India. DoorDash has launched a tasks app, and Maya notes force-torque and tactile data cannot come from humans.
- China builds state robot data factories. The government is funding centralised training centres where robots perform tasks while video and measurements are recorded. Priya argues thousands of New York homes give more random data than factory floors.
- Figure and Physical Intelligence split on models. Maya contrasts Figure's dual architecture, one model planning and one acting, with Physical Intelligence's single end-to-end pi-zero. That model emits about 100 candidate actions because actions have no average, and the industry is not aligned.
- Vercel cut 80% of its agent tools. A MongoDB paper found Vercel deleted 80% of its agent tools, lifting code success from 80% to 100% and cutting latency from 700 to 100 seconds. Maya says simpler harnesses often work better.
- Software engineering roles rise 14% year over year. Priya says software engineering roles are up 14% year over year, though big tech companies no longer hire straight out of college the way they once did.