On Sequoia's Training Data podcast, Box chief executive Aaron Levie explains why AI swept through software engineering and is moving far more slowly through legal work, sales and the rest of knowledge work: code is text, engineers fix their own broken connections, and their work already lives in GitHub. His conclusion is that the tedious work of getting AI into other people's workflows — not the models themselves — is where he is betting the money is.
A caller on the Moonshots AMA said he had cared for three loved ones for more than eight years and built a service, Tugboat Caregiving, for families like his. His problem: the people who most need to prepare for the next emergency are too tired to start. The panel's answer was to begin with the least impressive thing AI does — planning the day — and to work up from there.
On a Moonshots listener call, Matt from Terre Haute, Indiana asked what people with families, mortgages and a fixed location should do, since so much AI advice seems aimed at 22-year-olds working 80-hour weeks. The panel's answers: describe your actual obligations to a language model and brainstorm at the margin, practice on the tools at night, build a few online peer relationships, offer your domain knowledge to a marketplace like Mercor, and schedule slack time rather than waiting for spare time. The bigger forecasts about franchise-style openings from AI labs remain forecasts.
OpenAI's claimed Millennium Prize result used roughly 10,000 agents on a problem that was, as one entrepreneur on Moonshots put it, unusually easy to specify. The panel's argument: as the price of that kind of compute falls, the scarce skill becomes writing the target — and today's models, asked for ten ideas to cure cancer, produce a bad list.
A Moonshots panel unpacks Dwarkesh Patel and Jerry Han's experiment, which found that improvements in training data delivered a 12-fold compute-efficiency gain between 2019 and 2025 against 3.7-fold for architectures and training recipes — at small scale, on easy benchmarks. The panel then splits over whether a company's proprietary data is a durable advantage, with a $32 billion data-subsidiary valuation on one side and the fate of BloombergGPT on the other.
On Moonshots, the panel played a launch video for Redwood, an accelerator that Palo Alto startup Architect Labs says its AI designed end to end from a specification written by two architects. The company's paper reports two weeks to verified design and FPGA deployment, with a small language model running in a third week — while the headline 3.4-times efficiency figure comes from a projected Samsung 8-nanometer chip that has not been built. The panel, who disclosed they are investors and an advisor, argued the real story is a "designless" company and recursive self-improvement at the chip layer.
Philip Johnston spent six months failing to find a cheap trajectory to the nearest star system. A research campaign at the AI physics startup PSI, run on roughly 10 billion tokens and five or six hours of human time, returned an unintuitive answer: slow the spacecraft down first and let it fall toward the sun. The resulting Fermi Explorer mission proposes a 100-kilogram probe, a sub-$15 million budget, a launch by the end of 2029 and a journey of roughly 77,500 years.
An engineer who faked a missing editing feature using comment fields told Peregrine what to build next. A hurricane simulator stayed with one city. Co-founders Nick Noone and Ben Rudolph describe how they decide which piece of field improvisation becomes a product — and what they say it now costs to serve a city this way.
On The Diary of a CEO, David Friedberg argued that open-weight AI models will stop the industry's value from pooling in two or three labs, and wagered that someone with no money today will build a billion-dollar company on a model they downloaded. His case runs through the Netscape era, the fight in Washington over Chinese models, and a proposal that data centers generate their own power and sit in ordinary retirement accounts.
A Berkeley startup called Ditto has no feed and no swiping: students fill in a values questionnaire and a Wednesday evening text arrives with a match, a place and a time. On Moonshots, the panel weighed the appeal of having the choosing removed against Emad's worry about outsourcing connection, Dave's warning that a system that knows you well is also a very good salesperson, and Alex's refusal to use it at all.
On Moonshots EP #283, Peter Diamandis introduced a reported $6 billion NVIDIA arrangement with the coding startup Poolside as America's answer to Chinese open models. Emad Mostaque argued the real driver is selling more GPUs, while Alex and Dave disagreed about whether licensing-and-hiring deals exist to dodge antitrust review or simply to hire fast — and what happens to the half of Poolside that stays behind.
On Training Data, Parag Agrawal explains how his company Parallel entered web search without first building a giant index: it launched a search agent that crawled after a request arrived, replaced outsourced human data collection for insurance, sales and finance customers, and treated the index as a latency optimization to be grown later. He describes the agent-specific architecture behind it, the 200-millisecond Turbo mode Parallel announced in July, and a Google Cloud deal that puts Parallel Search beside Google Search as a grounding option.