AI Builders Digest

In five years, 90% of enterprise tokens will come from work nobody asked for

The bet that one or two labs capture 95% of AI value is the one worth fading. Models being smart is not the bottleneck: real workflows need data connections, human-in-the-loop moments, idle agents waiting on delays, change management, and legacy systems nobody modernized, and no research organization wants to attack that list. Token subsidization from the labs is temporary because public markets will eventually apply the same laws of capitalism to everyone, and non-economic actors like Meta, China, and NVIDIA will happily run inference at 10% margin, which pushes value down to the application layer either way. The sharpest prediction: "in five years from now, I would bet, like, 90% of all tokens in the enterprise are things that a user never kicked off, and they just see a result." On why coding diffused instantly and nothing else has: code's entire value is text a person types at a computer, the audience debugs its own MCP errors instead of calling IT, and there is no equivalent of "just connect your GitHub" for knowledge work.

• #agents
• #enterprise
• #open-source

A model that outputs probabilities, not words, at 600x lower cost than a frontier judge

Every almost never tests new foundation models and has been running this one for about a week. It doesn't produce words as output, it produces probabilities, which lets it stand in as a judge in cases that would otherwise need a Fable-level model. In testing it came out 25x faster and 600x cheaper. His framing: this is the kind of thing that looks obviously indispensable 6 to 12 months from now.

• #evals
• #products

Better models make the applied AI layer more necessary, not less

There's a massive chasm between what models can do and the workflows enterprises actually want automated, and filling it means connecting intelligence to workflows, reengineering processes, aggregating context, building human-in-the-loop moments, running domain specific evals, and managing security and governance. The counterintuitive part is that this layer gets more important as models improve, not less: greater capability means more complex tasks get attempted, which amplifies the damage when the connective work is done badly. He expects this layer to emerge in every vertical and horizontal category.

• #agents
• #enterprise

Profound raises $180M Series D at $1.8B with a third of the Fortune 100 on board

The round is co-led by Sequoia and Kleiner Perkins, with Lightspeed, Khosla, Saga, Evantic, and SPC participating. The origin story is the interesting part: the two founders met at SPC, an Uber maps engineer and a founder with no college degree, and neither showed up with the idea. The partnership formed first, then the company. "We didn't find the company. We watched it form."

• #funding
• #products

Vercel Labs goes public, and hiding the model choice hurts customers

Vercel Labs is now the formal home for the company's in-public research and experimentation, including what didn't pan out. Separately, a strong position on model abstraction: the future is multi-model, and trying to hide the choice confuses customers and blocks them from participating in the upside of the most exciting market competition of our times, or from mastering the best tool for a given job. Also flagged Safari 27 shipping JSPI, which lets synchronous native WebAssembly code suspend on an async Promise, with an expectation that WebAssembly plays a huge role as more code goes native.

• #products
• #open-source

MCPs now beat CLIs for most integrations, against expectation

A reversal he says he wasn't expecting. Models have gotten much better at tool calling, tools can be deferred, and MCP is now stateless, which removes the reasons CLIs used to win. The practical advice for anyone building: if you need to compose or filter data, add parameters like query to your MCP tools rather than reaching for a CLI.

• #agents
• #open-source

Salesforce in Claude hits beta with 37 pre-built sales skills

Accounts, opportunities, and pipeline now live inside Claude. The 37 pre-built sales skills cover prepping a call, reviewing a deal, building a pipeline dashboard, and sending a forecast, all without leaving the conversation.

• #products
• #enterprise

Claude for Small Business adds 43 workflows and 27 integrations after 900,000 installs

New connectors include Shopify, Salesforce, TikTok, Atlassian, Zoom, Xero, Gusto, Square, Stripe, and Zapier. The release is built on what owners asked for on the spring tour, where about a third of the 1,000-plus owners across 10 cities wanted help growing the business rather than running the back office: generating leads, answering inbound inquiries, and writing proposals. The tour returns this fall with free workshops in 10 US cities, over 750 community workshops from 150-plus approved trainers, and free webinars from 14 integration partners. One owner reports making $20,000 in six weeks on proposals written with Claude.

• #products
• #enterprise

Gemini Notebook gets spoken Q&A in ~100 languages and auto-saving lecture recordings

Students can now hold a live, spoken back-and-forth with their own class materials in roughly 100 languages. The second addition is on-the-go recording of lectures and notes, with the audio saving automatically into the notebook you picked. University students can still get a free Google AI Plan in 140-plus countries for higher limits and access to more Google products.

• #products
• #education

A harness layer cut agent fix waves to half the time on identical frontier models

Running capy.ai with GStack and GBrain on a batch of outstanding issues and PRs took about half the time that raw Codex or Claude Code would have needed for a day of work, using the same underlying frontier models. His word for it was amazed. He also called Muse the likely winner in a separate exchange.

• #agents
• #products

Too many founders assume the next round happens no matter what

Capital is a weapon for acceleration, and used well it makes you unbeatable, but dependence on it turns an ordinary downturn into something much worse. The advice he's giving portfolio founders is to identify the default path that guarantees the company survives as a self-sustaining business, then separately model what abundant capital and what no capital each do to that business. Nobody knows how capital markets evolve over the next 6 to 18 months.

• #funding

The AI naming curse: safety firms making AI unsafe, effective altruists being neither

A blunt shot at how the industry names itself, running through an "AI Safety" firm that made AI unsafe, effective altruists he calls both ineffective and enabling criminal activity, and a firm named Irregular that he says is regularly incompetent. On the technical side, a pointed question about probability-emitting models: if your output domain is known in advance, why not just train a model to produce logprobs over enums?

• #policy
• #evals

Design the solopreneur business around work you enjoy, then delegate the rest to bots

Three months into doing it full time, his advice is that there will always be choices that make more money, but if those choices leave you doing work you don't enjoy, you've defeated the point of going solo in the first place. Delegate the boring work to bots or just don't do it. He also collected the standout agent use cases people sent in: monitoring hundreds of ad accounts to catch broken signup and payment flows, watching Jira to flag blockers and draft the weekly status update, triaging support tickets with escalation context, and regularly testing a hotel's booking flow so website errors get caught.

• #agents
• #products