Everyone agreed memory is the rate limiter. Almost nobody said which memory they meant. Builders heard context windows. A week earlier, SpaceX and Tesla had committed $16.8 billion to a fab that makes its own logic and memory. Those are the same constraint seen from opposite ends. Generating a token is mostly a read, not a calculation. The model streams its weights and its accumulated context out of memory for every token it produces. An agent makes that worse on purpose. It keeps state, re-reads it every step, and takes hundreds of steps where a chat takes one. Add FLOPs to that and very little happens. The accelerator sits waiting on the memory bus. So agent memory and high bandwidth memory are not two subjects. Agent memory is the workload, HBM is the pipe, and the pipe is the scarce part. Three companies make over 95% of the world's DRAM, and every wafer they move to HBM is a wafer that never becomes the RAM in a laptop. TrendForce has conventional DRAM contract prices rising 58 to 63% this quarter. J.P. Morgan puts the full move above 400% between the start of 2024 and the end of 2026, and expects the shortage to run for years. That is what the rate limiter really is. Not a flaw in anyone's agent framework, but an allocation decision made by three suppliers that then reprices hardware for everyone standing downstream. Which is why the response taking shape is a building rather than a framework. When the bottleneck moves from compute to memory, leverage moves from whoever trains the best model to whoever decides where the wafers go. A fab is the only way out of being a customer of that decision.

Anthropic is now the only frontier lab that has not signed the open weights letter. OpenAI signed within hours of Jensen's post. Google signed today through Sundar. The consensus take says Anthropic caves next, because being the lone holdout in the middle of a Washington fight looks terrible. I think the consensus has this backwards. Anthropic is the one lab that cannot sign cheaply. OpenAI ships open weight models of its own. Google ships Gemma. Their signatures cost nothing. Anthropic has never released open weights, reportedly lobbied for the Chinese open weight restrictions this letter opposes, and just spent 19 days in June with its flagship suspended by an export order. Signing would reverse an active lobbying position and erase the one thing that makes them different. Their new pricing already monetizes the closed position. The risk tier carries the premium. So my call is that Anthropic does not sign. What we get instead is an essay. Something long and thoughtful about how open and closed both matter, with a warning about frontier weights, and no signature under the letter. If they sign within 2 weeks, the pressure beat the incentives and I was wrong. Logged either way.

Semantic search has no concept of time. Save a better version of an old note and cosine similarity happily returns the stale one. So ranking in my second brain decays on purpose. Every result gets a small recency boost that fades out over about 90 days, capped low enough that meaning always wins. A fresh item beats a stale near duplicate in a near tie. It cannot beat a genuinely better match. I stopped filing and tagging entirely. The decay does that job now.

I tested six image models on the same tee before picking one for GhostModel. Same wrinkled flat lay in, same instruction for every model. Keep the print, the colors, and the artist's signature exactly as they are. The $0.08 model turned my t shirt into a sweatshirt. The most famous one was faithful but took 87 seconds a shot. One kept the print perfect and then invented a line of text that was never on the shirt. The winner was the cheapest model in the field. Seedream 4.5, 4 cents a shot. It held the print, the signature, and the true black fabric across all four scenes. That model now runs GhostModel. And the studio is open. One photo of your garment in, your scenes out. 8 free generations, no card, just an email. Link below. The whole bench cost 39 cents. Guessing would have been free and wrong.

One phone photo. Four photoshoots. 60 seconds, about $0.30. This morning I clicked picture of a wrinkled flat lay of a tee on my dining table. Tonight, the same tee on a model in a studio, on a night street, at golden hour, at a cafe. The artist's signature on the print survived all four scenes. One miss was that the night-street scene redrew the tiger's face slightly. 3 of 4 were clean. That is the actual state of image models in July 2026. Cost Per scene for ~15 seconds, ~$0.05. A four location shoot with a booked model runs into the thousands. The garment is the constant. Everything else is a prompt.

Everyone spent yesterday arguing whether Kimi K3 got 100% faster after Moonshot sold out of subscriptions. I had a cleaner option re-run my bench from yesterday. Same landing page spec, same 12 acceptance criteria, same harness, same model. One shot, no coaching, both times. Yesterday's run a complete, deployable page. 12 for 12, in 35 minutes. This morning's run: 26 minutes, and it never built the page. It wrote a config file, drew two placeholder SVGs of a sweater, and exited. The homepage still says "To get started, edit page.tsx." So I cannot tell you whether K3 got faster. The run finished 26% quicker while doing roughly none of the work. What I can tell you is worth more. A one shot agent run is a draw from a distribution. Yesterday I posted "12/12 first shot." Today's run put error bars on my own headline. The lesson for anyone building with agents is that acceptance criteria plus a retry budget, every time. One shot and trust is not a workflow. It is a slot machine.

Tried to verify the speedup the direct way. So I re ran Yesterday's landing page bench this morning. Same spec, same harness, same model. Got an answer to a different question! Previous's run shipped a complete page, 12/12 on my acceptance criteria, in 35 minutes. Today's run spent 26 minutes drawing two placeholder SVGs and exited without ever writing the page! So no speed verdict from me, n=2. But one shot variance at this scale is a bigger operational cost than tokens per second, and not a lot of people are pricing it.

Kimi K3 is #1 on LMArena's frontend leaderboard at a third of Claude's price. I gave both the same task to build the landing page for a product I'm shipping this month. Same spec, 12 acceptance criteria, written like a PRD. One shot each, no coaching. Both scored 12 out of 12. First try, zero fixes. On capability, the hype is real. Everything else was the story. Claude finished in 7 minutes. K3 took 35, because it ran its own browser QA loop, screenshotting desktop and mobile like a very thorough intern. The bill for all that diligence: 2.8 million tokens, 52 requests, $1.88. Claude's run metered 75k tokens and rode the Max subscription I already pay for. At Fable 5 API prices, K3's exact workload would have been $6.27. Then the spooky part was something with zero shared context, Claude wrote "One flat lay in. A full shoot out." K3 wrote "One flat lay in. An entire shoot out." Their error messages match word for word. One difference is that K3 invented product commitments I never made like 25 free credits, essentially locking the founder pricing for life. Claude kept the promises exactly as vague as the spec did. For a PM building solo, the cheap model ships the same page for the price of a coffee. It just spends more of your afternoon, and you read its copy like a lawyer before it signs contracts on your behalf.

My agents can browse the web, read docs, and write code. They knew nothing about what I know. One endpoint fixed it. My knowledge base is now three tools any AI conversation can call: search it, ask it, save to it. The whole system: two Postgres tables, zero plugins, a week of evenings to build. Nothing runs when I am not using it. Why second brains die, and the five design decisions that fixed mine: article below.