AI Builders Digest

Price per token is a broken unit of comparison across models

Two pizzas, one cut into 8 slices at $2 and one into 16 at $1.25: the cheaper slice costs you more per pizza. Same with tokenizers. In one comparison across English, technical, multilingual, and numerical text, GPT-5.6 Sol's tokenizer used 766 tokens where Claude Opus 5 was estimated at 1,170, roughly 34.5% fewer for identical text. The real metric is price per successful outcome, which means benchmarks are a starting point and your own use case is the actual test.

• #pricing
• #evals

Scaling laws are empirical observations, not laws of physics

The claim that AI structurally centralizes power because it is compute hungry ignores 125 years of super exponential improvement in compute price-performance. Scaling laws hold for particular architectures, objectives, and datasets; change any one and the curve changes. If the brain is any guide, real AGI will be very efficient, which makes today's scaling behavior a symptom of how inefficient current ML algorithms are rather than a permanent constraint requiring a datacenter.

• #policy
• #hardware

The next consumer AI winner will be multiplayer, not a better model

Consumer platforms move along a slider from deeply technical founders to product geniuses as the underlying tech matures: Google was roughly 95% backend magic, Facebook less so, and by Pinterest, Snap, and Instagram the CEOs were not technical at all. AI is still at the Google end, with ChatGPT essentially a text box and Character AI a product that was really the model. Custom GPTs are the tell: "it's clearly made by a team that is unbelievably capable but isn't social." There is no visibility into what prompt or documents sit behind one, so no trust and no status-seeking loop, which is what makes a community actually work. The open question is whether incumbent gravity, especially memory and habit, is simply too strong for a new social layer to pull power users away.

• #products
• #agents

shadcn is what people actually wanted from React

React was the spec, the description of the geometry of the bricks, not the LEGO bricks themselves. shadcn is the part people wanted: reusable, high quality, tunable components. Its shape matters for the AI era specifically, since it is a pseudo-library where the code exists to be pulled into a context window and remixed rather than imported and left alone.

• #open-source
• #products

The bureaucracy blocking your deal is the moat you are building

A stealth founder in an unsexy, deeply fragmented market is stuck at the one-yard line on contracts that would bring credibility, revenue, and network effects. Champions and terms are aligned; it is summer and one random clause. The right reframe is how long it would take anyone else to replicate that position, even with existing leverage and relationships. Performative 996 suffering is not a moat, but entrenched software and relationships that take real pain to build are.

• #startups

B2B software no longer has an excuse for bad UX

AI removes the last justification for enterprise tools being harder to use than consumer ones. Every software product can and should now meet the bar set by the best consumer apps.

• #products

The AI illuminati group chats mostly do not exist

Top-tier people in tech meet and know each other far less often than outsiders assume. Secret group chats exist but tend to be short lived exceptions. What that implies is that the major headlines are roughly what insiders see too, and that most effort still goes into doing the work rather than working the narrative.

• #culture

An interactive explainer for how lossless watermarking works

Watermarking without quality loss is counterintuitive and does not feel like it should work at all. Rather than take it on faith, the fix was building an artifact with Claude to walk through the mechanism, now shared publicly.

• #research

Grok Bot cannot reach the one data source that would differentiate it

X is the single most differentiated data source Grok Bot could have, and the connector does not work; logging into X from the cloud computer fails too. Separately, an endorsement of Dario's position that using AI to cure disease, plus speeding regulatory approval of healthcare breakthroughs, could deliver 10x the benefit of everything else combined. A forthcoming episode with Riley Brown shows Codex pulling 100 top-performing thumbnails in a niche into a Paper canvas, then remixing them with his own photos.

• #products
• #agents

Record crop production is what abundance actually looks like

Against the doom on most timelines, global crop production keeps setting records year on year, up and to the right. That curve is the output of markets and technology compounding together over millennia, not a recent accident.

• #policy