Tokens Go Gray-Market, the Laptop's Last Six Months & Prompt Engineering Becomes Goal Engineering
The Frontier Goes Gray-Market
If yesterday was the day the frontier went behind glass, today was the day the side door opened. The access story stopped being a lament and became a logistics story — about how, exactly, the rest of the world keeps buying the models it's no longer allowed to have.
The field guide. Peter Steinberger (60 replies, 77 RTs, 874 likes, 73K views) shared a ChinaTalk piece — "How to Buy Cheap Claude Tokens in China" — with the only quote it needed: "History teaches us that access blockage rarely stops determined users." Pressed on when normal users get Sol, he stayed out of it: "I build cool things with AI, I don't decide politics." The replies filled in the mechanics. Vishal Singh: "It's just insane how Chinese students are buying GPT-5.4/5.5 and Claude API access from Xianyu/Taobao proxy sellers for almost 96–97% cheaper, while they promote Open Source AI to the rest of the world." Silverrock put it in historical company: "Prohibition, the Great Firewall, software piracy… determined users almost always route around access blocks. The transfer station economy is just the latest, highly optimized version." Ram Bakthavachalam: "movie piracy equivalent of llm." And the line of the day, from Joe Doliner (13 likes): "I've got a peptide guy and I've got a token guy. (It's the same guy.)"
The policy got dumber. The strongest substantive thread came from LLMJunky (12 replies, 4 RTs, 136 likes, 8.2K views): "The USG really asked OpenAI to withhold 5.6 Terra and 5.6 Luna from us, despite the FACT they are either the same or worse than the models we already have access to. Except at a huge discount. So not only do we not get Sol, but we have to continue to pay >2x for the same intelligence. … Wanting to hold back Sol at least has some justification. The White House has absolutely no justification to withhold the other two models." The reply that landed hardest was the FAFO read. Anthony Poschen (3 likes): "you can't claim you created something as capable of changing society as a nuke (yes they did this) and not expect serious government to feel obligated to preemptively evaluate your claimed nukes… This is what the big AI labs wanted. They got what they asked for." LLMJunky's reply was the whole essay in nine words: "Yeah if they were nukes why the hell did you build them." The mood on timing, from Matt Banham: "we'll be lucky if we see anything before August at least."
The satire wrote itself. LLMJunky (12 replies, 71 likes, 4K views): "This application to join the GPT 5.6 Sol preview is wild" — the replies imagined the form as a loyalty oath, capped by Astrohacker: "i pledge allegiance to the flag of the united states of ai… one nation, under the superintelligent gpu cluster, indivisible, with censorship-free models and unlimited tokens for everybody." Theo distilled it into a quote-tweet (of Doliner again) — "This is a banger but also kind of true" — over the verdict: "The new benchmark is how long the government bans you for."
The self-reliance impulse. Underneath the jokes, the same instinct kept surfacing: build the off-ramp yourself. LLMJunky: "I want to start a potluck but instead of food, we all bring our GPUs. … Solar powered. Train, fine tune, run, and control our own models. I don't want the government or company deciding I'm too dangerous to be allowed access to intelligence. … We're going to have to take it into our own hands." And the open-weights tier kept making that easier — WeZZard: "Good news: GLM 5.2 is better than Claude in real use now, except the vision support. Bad news: GLM 5.2's mother company is heavily constrained with GPU powers." The through-line is the one that's defined the week: a frontier that pulls away from its builders doesn't stop them — it just routes them somewhere uglier, cheaper, and harder to govern.
Six Months to Agents Off the Laptop
The day's biggest builder thread had nothing to do with policy and everything to do with where your agent actually runs. Theo (270 replies, 22 RTs, 2,043 likes, 257K views): "I'd estimate we're ~6 months from most devs moving their code agents off of their laptops." It's the natural extrapolation of this week's T3 Code remote-first push — and the replies turned it into the most useful infrastructure argument in a while.
The "it'll get better" camp. Rhys Sullivan: "if the tooling gets better yeah otherwise i dont see it" — Theo: "It will necessarily get better." The "this is already how it works" reports came in fast: Samir Chowdhury: "This is how it was at Amazon. All devs had a cloud instance to develop against… When coding agents came, I'd run multiple on my cloud instance." Panos Zepos: "T3 code running on an oracle free-for-life vm is SUPER close. Like close enough to make a MacBook Neo an attractive option.. all you need is safari."
The "primitives are wrong" camp had the sharpest critique of all. Isaac Dobbins: "right now the primitives are all wrong. Tailscale helps… but I also feel like SSH is not the right layer to build this all on. … The idealistic extreme here is that the boundary between computers is completely dissolved. I want to run grep from my terminal and have that executed on any computer I desire. Better yet I want my agent to figure out which computer to be running tool calls on and I never want to think about external execution." Joshua Scarsbrook named the dream directly: a Single System Image, "A whole cluster of local machines, virtual machines, and cloud machines all appearing as the same system with a unified interface." On the gaps, addison wanted "a far better way to track changes / sync between devices / handle networking between remote & local" — Theo: "Working on it!"
The backlash was loud. Theo amplified the strongest counter himself, RT'ing maria with "i don't know what to think anymore" over her contrarian call (106 likes in-thread): "all these agent coding TUIs are a phase and it will be short lived. Most devs will be back in GUIs and IDEs in a few months." Zec Fingerhut: "I will probably never move off of my laptop… i dont wanna look at tiny ass git diffs from my phone." And Aben pointed the other way entirely: "local models and devices that can run those models would [be] the next frontier. Can't trust big tech with such dependencies." The platform crowd, meanwhile, just linked their products — coder/coder, namespace devboxes — and the iOS devs grieved: Thomas Ricouard, "iOS dev in shamble[s]," to which Theo could only agree: "…yeah it's rough." The split is real, but notice the shape: even the skeptics are arguing about which remote primitive wins, not whether the work leaves the laptop.
From Prompt Engineering to Goal & Eval Engineering
While the timeline argued about where agents run, Jerry Liu made the sharper claim about how we drive them. Jerry Liu (31 replies, 6 RTs, 135 likes, 13.5K views), from playing with the /goal primitive: "there's less and less of a need to build any type of workflow manually (whether through code, drag and drop, or a prompt). Instead, specify the goal, let the model intelligence figure out the underlying steps. If the task is repeatable, then you can gather a dataset with ground-truth, and hillclimb it… The world is moving from prompt engineering → goal and eval engineering."
The diagnosis resonated — with caveats. isaac (4 likes): "Doing this is what dspy is all about!" Jonny Gravity: "The pay-off of /goal is simply that it orchestrates the execute→verify→continue loop for you. Where we come in is ensuring that the /goal converts non-deterministic execution into deterministic results." The most concrete field report came from Jordan Hochenbaum: "I would recommend wrapping /goal in a workflow ('ultra code') in Claude Code as a power move… the most important part, the verification subloop. We've had a lot of success applying this to large scale migrations (e.g. converting a bunch of legacy frontend code from Ember to React), performance optimization for CI, and writing/maintaining doc evals that cut the time it takes agents to find the right information by 50%." The skeptics aimed at the same soft spot — the eval. Michał Piszczek: "goal + eval engineering is clean when ground truth exists. the messy half of prod is the non-repeatable tasks with no eval set, just an owner holding the bag." Anders Lie: "you usually want pretty well defined completion or optimization criteria, which is hard to enforce… provides no strong guarantees yet." And Rorschach supplied the inevitable meme: "LOOP ENGINEERING! ENGINEERING LOOPS THAT ENGINEER LOOPS."
The fight is now a literal main-stage event. That exact tension — autonomy vs. verifiable correctness — has been promoted from timeline argument to scheduled bout. Allie Howe, RT'd by swyx: "THE GREAT LOOPS DEBATE is on. @ianlivingstone and @GeoffreyHuntley take on @dexhorthy and a surprise guest to answer: Do loops produce slop? Do we have the AI infra we need today to verify the work and track state? What tasks are right for loops? Main stage 7/2 @aiDotEngineer." The replies pre-staked the two poles cleanly — Ben Freed: "Loops prevent slop," against 0xBender's framing that "slop vs real verification is the fight everyone building agents needs." If 2025 was prompt-crafting and early 2026 was loop-crafting, the throughline of today is that the grader — the eval, the verifier, the ground truth — is becoming the actual job.
The Prep Stack & the /research Debate
Matt Pocock spent the weekend formalizing the part of the work that precedes the loop. Pocock (42 replies, 19 RTs, 468 likes, 35K views): "The 3 main 'prep' activities in software development are: Discuss (figure out what to build), Research (summarize world knowledge to aid discussion), Prototype (build something to aid discussion). First question: did I miss any? Second question: how was it not obvious that AI would be really fucking helpful here."
His own pipeline is a chain of skills. Pocock: "/grill-with-docs → /handoff → /prototype → /handoff back to original session." The replies extended the taxonomy — Wogan May proposed a fourth verb, "invalidate: figure out why an approach won't work asap and document it so you don't go down rabbit holes," and Leigh Stillard added "'codify' … and then 'break' with… codex — an adversarial review, and the high confidence challenges are automatically adopted." The sharpest pushback questioned the whole framing — Adel Bucetta: "our industry hasn't solved the 'what to build' problem, we just outsourced it to ai and called it prep activities" — to which Pocock had the best line: "You can't solve the 'what to build' problem any more than an artist can solve the 'what to paint' problem."
Should /research be a skill? By morning, the demand had crystallized into a real design question. Pocock (21 replies, 65 likes, 6.8K views): "Folks are asking me to make a /research skill, but I'm not sure. The skill would: run a research agent in the background, save its results to a .md file, tell it to continue its original activity (probably grilling) if possible. I tend to just say that verbatim — which tells me it probably should be a skill." The best replies were about the output contract, not the search. Timur Yessenov: "I'd make it a skill only if it forces a tiny finished file: question, sources checked, what changed my mind, open questions, and where the original task should resume. Background research without that handoff just becomes another invisible tab." Joakim Hauge: "The interesting part isn't the research, it's making sure the research agent knows when enough is enough." (Cross-harness logistics came up too — Pocock: "Most harnesses use .agents/skills… So I symlink them to .agents/skills.") It's a small thread, but it's a clean snapshot of how the skills ecosystem actually matures: a phrase you keep typing verbatim becomes a skill, and the value is in the file it's forced to leave behind.
Also Worth a Look
AI Engineer World's Fair sells out. The biggest in-person AI event of the year (June 29 – July 2, SF; 29 tracks, 300 speakers, 6,000+ attendees) crossed into sold-out territory — swyx: "we crossed our 6k attendee mark a while ago. will probably call sold out when we hit 7k this weekend… this is the epicenter of ai next week." His own track is the tell of the moment: swyx: "we have been scaling without slop by working with aligned domain experts… it's clear that FDE [Forward Deployed Engineering] is one of the most in demand disciplines on earth," debuting a first-ever AI FDE miniconference. Expect next week's roundups to be heavy on dispatches.
Codex ships another QoL batch. Quiet but concrete — OpenAI Developers (via reach_vb): "another batch of Codex quality-of-life improvements this week: smoother long threads with position preservation and jump navigation, better Settings search, zoom-safe UI positioning, faster thread switching and deeper local history." The unglamorous interface work that long-running agent sessions actually need.
LiteParse goes open source. Jerry Liu's parsing tear continued — Liu: "LiteParse is unreasonably good for document parsing… average parse time per page is 3ms… it tops opendataloader-bench, OlmOCR-bench, and ParseBench in terms of accuracy… It even gives you basic bounding boxes that your coding agent can stitch together" — now open source and past 10K GitHub stars, positioned as a fast first pass before you reach for VLM-grade LlamaParse.
Report dollars, not tokens. A nerdier note from swyx worth filing: taking Noam Brown's constant-inference-budget rule seriously means "open models have a lot more dollar-per-token mileage than closed model APIs. So anyone launching an open model… should report thinking levels measured by dollar inference on popular inference providers, instead of by number of tokens on the x axis." A small methodological shift that would make a lot of "we beat the frontier" charts look very different.
The frontier was built worldwide. A pointed counter to the access-hoarding mood, antirez (RT'd by Armin Ronacher): "Modern AI resulted from research made also by many non-US scientists (Hinton, the French folks, Linnainmaa, many others). The pre-training corpus was produced worldwide with massive code contribution from Europe OSS. What is happening with frontier LLMs is unacceptable."
Agents and software, unbundled. Jerry Liu, on the "is SaaS dead" discourse — Liu: "there's a lot of value in both agents and software (I don't think SaaS is dead), but the product interfaces are different… the interfaces will unbundle between the two over time." Phi Browser had the vivid version: "today [your tool] only has a human-shaped interface, so I drive it through your UI pixel by pixel like a clumsy API. The unbundle you're predicting is the tool finally growing a face meant for me too."