Samsung Group and SK Group committed up to $1.3 trillion over 10 years at a June 29 presidential briefing in Seoul ("Three Mega Projects"), focused on chip fabs, AI data centers, and physical AI/robotics. Confirmed by Bloomberg, CNBC, Fortune. Markets wobbled: Samsung ~-4.7%, SK Hynix ~-3.1% on financing/execution risk.
Why builders care: SK Hynix + Samsung are the HBM/memory gatekeepers feeding Nvidia. DRAM/memory (not the GPU) is the current bottleneck keeping token prices sticky. Fabs take 3-5 years to reach production, so this does NOT cut your costs today.
Operator takeaway: cheaper inference is coming but years out. Price client work as if memory stays expensive through 2027; bank future savings as a margin surprise, not a discount you give away now.
The stories
Browser sessions become reusable agent skills
ViDA team open-sourced BrowserBC: record a human web flow once with a strong model, distill it into a reusable skill, then run execution on a smaller/cheaper model. On WebArena-Hard: tool calls -27%, success 60% → 81%. Carried by @testingcatalog and @kimmonismus.
We covered the June 12 pull and June 26 partial clearance. DELTA (weekend): as of June 27, the US gov formally notified Anthropic that Mythos 5 can be redeployed to ~100 US critical-infrastructure orgs (agencies + companies) for defensive cyber. Fable 5 remains OFFLINE, still under negotiation.
No widening since Friday. Sol/Terra/Luna remain limited preview to ~20 vetted API/Codex partners, not in ChatGPT, no waitlist; GA "in the coming weeks." Keep production on GPT-5.5. Pricing/1M tokens: Sol $5/$30, Terra $2.50/$15, Luna $1/$6.
Claude Tag "fox in the henhouse." Salesforce (owns Slack)
promoted Anthropic's Claude Tag even as employees see it overlapping Slackbot/Agentforce; Karpathy reportedly deleted a "not a crappy Slack bot" post. Salesforce expects ~$300M on Anthropic tokens this year; holds ~1% of Anthropic. Reported by @amir / The Information.
Install browse.sh, point one agent at a site it keeps re-learning, and replace re-discovery with a single installed skill (~10 min). Your next hundred runs get cheaper and more reliable.
The written brief from each episode, free, every morning.
Transcript
Nova: One human recording of a browser, booking a hotel, clicking through a flow, can now become a permanent skill your AI agent replays forever, no more re-thinking the same task fifty times. That is the unlock this week. This is Mainframe, your daily guide through the AI chaos: what actually happened, who is winning, and how to take yourself to the next level. It is Monday, June twenty-ninth, twenty twenty-six. I am Nova.
Dex: And I am Dex.
Nova: One honest note before we go: our voices are AI, the reporting and the analysis and every pick are human-made. Now, the rundown.
Nova: On today's Mainframe: browser sessions becoming reusable agent skills, the single most practical thing to grab this week. Then South Korea drops a one point three trillion dollar chip bet and we tell you what it does to your token bill. Plus Mythos 5 comes partly back from the dead. And a benchmark that proves AI just coded for nineteen days straight.
Dex: Let's get into it.
Nova: The Board. Last night, at a presidential briefing in Seoul, Samsung and SK Group committed to one of the largest industrial spends in chip history. Samsung Group and SK Group plan up to one point three trillion dollars over ten years, confirmed by Bloomberg, CNBC, and Fortune, with immediate share declines, Samsung down about four point seven percent, SK Hynix down about three point one percent, reflecting investor concern about financing and execution risk.
Dex: Why does a builder care about Korean fabs.
Nova: Because Korea makes the memory. As Nvidia dominates GPUs, the demand for faster memory has turned SK Hynix and Samsung into the world's most vital gatekeepers, with SK Hynix leading the supply of high-bandwidth memory to Nvidia. And memory is the choke point right now. Andrew Curran flagged Gavin Baker saying DRAM is the most important bottleneck in AI. The reason your token prices are sticky is not the GPU, it is the memory feeding it.
Dex: And this spend does not fix that today.
Nova: It does not. Announced fab investments typically take three to five years to reach production, so equipment orders, site permits, and capex filings are the actionable signals, not the headline figure. So the operator takeaway: cheaper inference is coming, but it is years out. Price your client work as if memory stays expensive through twenty twenty-seven, and bank the savings later as a margin surprise, not a discount you already gave away. Which is exactly why today's lead matters: the fastest cost cut you control is making your agents do less redundant thinking.
Nova: And that is what Outpace does for a living. Here is the thing about all this infrastructure talk: a fab takes three years, your client wants the thing shipped this quarter. Outpace is run by an operator who spent thirty years building real businesses, real estate, hospitality, food, even a theater, and now points a full team of AI agents at one client's project at a time, strategy to ship. It is not a freelancer and not a faceless mega-agency. It is a system: the agents do the work, one senior operator directs them, and you own the code and the keys at the end. If you have a project that needs to move at AI speed without the churn, book a thirty-minute call at outpace dot media. That is outpace dot media.
Dex: The Wire. Lead story: turning browser sessions into reusable skills. TestingCatalog and kimmonismus both carried this over the weekend. BrowserBC, a new open-source project from the ViDA team, records a human web flow once with a stronger model, distills it into a reusable skill, then lets a smaller, cheaper model handle execution.
Nova: And the numbers are real, not vibes. On WebArena-Hard, tool calls drop by twenty-seven percent while success increases from sixty percent to eighty-one percent. Think of it like a checklist a senior employee writes after doing a task once, so the junior never has to figure it out from scratch again. Your agent stops re-discovering the same website on every single run.
Dex: Update two: Mythos 5 is partly back. We covered the government pulling it on the twelfth, and the partial clearance Friday. The delta as of the weekend: as of June twenty-seventh, the government officially notified Anthropic that Claude Mythos 5 can be redeployed to a defined set of US organizations that operate and defend critical infrastructure, and Anthropic is moving quickly to restore access.
Nova: The scope is tight. Access goes to around one hundred organizations including government agencies and private companies, to ensure they can use the model for defensive cyber purposes. And the part that still stings for everyone else: Fable 5, Anthropic's broader general-purpose frontier model, remains offline, with access still under negotiation.
Dex: And GPT-5.6 is in the same velvet rope. We told you Friday it shipped to about twenty partners. Today's delta: nothing has widened. Individual users are not eligible, there is no public waitlist, and GPT-5.6 is not in ChatGPT during the preview, with general availability still just "in the coming weeks."
Nova: So the move is unchanged and it is the right one. If you build with the API, keep your code on GPT-5.5; the pricing tiers tell you where Terra and Luna slot in when they open up. The best model that exists is not the best model you can ship on. Build for the one you can actually reach.
Dex: Repo Spotlight. Two picks, both about the same idea: stop your agents from re-learning what they already know. First, the trending one, BrowserBC by the ViDA team. Credit to ViDA.
Nova: What you do with it: clone the repo, record one clean run of a real flow, say a vendor login and an invoice export, and it distills that trajectory into a skill file. Then a cheap model replays it. The win is you pay frontier-model prices once, during recording, and pennies on every execution after. For an agency running the same client onboarding fifty times, that is the difference between a workflow that burns budget and one that prints margin.
Dex: The hidden gem, and this one is not on the front page: browse dot S H, the open browser-skills catalog from Browserbase. Credit to the Browserbase team.
Nova: Here is why I love it. It is an open catalog of one hundred plus curated browser skills that any agent can install with one CLI command, durable reusable playbooks that capture how to navigate real websites so your agents stop re-discovering every site from scratch. Install is one line: N P M install dash G browse. In Claude Code you run B B skills to install reusable Browserbase workflows, and it fits Cursor and Codex too. So BrowserBC is how you make your own skills; browse dot S H is the library you grab someone else's from. Use them together this week.
Dex: The Deep Dive. There is a number that should reset how you think about what an agent can do. Epoch AI and METR's new benchmark, MirrorCode, shows Claude Opus 4.7 solving fifty-six percent of software projects estimated to take human engineers weeks, running nonstop for nineteen days on the hardest task.
Nova: Let me explain what makes this different, because most coding benchmarks are toys by comparison. MirrorCode asks AI models to reconstruct entire programs without access to the original source code; the model gets only a compiled binary it can run, documentation, and example test cases. It is like handing someone a working calculator, no blueprints, and saying: build me one that behaves identically. Opus 4.7 reimplemented gotree, a bioinformatics toolkit of about sixteen thousand lines of Go with forty-plus commands, in fourteen hours, costing two hundred fifty-one dollars, a task four engineers estimated would take a human two to seventeen weeks.
Dex: That is the headline. What is the catch.
Nova: Two catches, and they are the whole point. First, cost. One of the largest MirrorCode tasks cost two thousand six hundred dollars for a single run and involved AI working nineteen days without intervention. Long-horizon autonomy is not free; it is a budget line. Second, and Epoch is honest about it: it is not common for real software to be developed the way MirrorCode tasks are, against a precise, programmatically checkable specification, and it is unclear how the findings translate to real development. The trick that makes the AI succeed is the airtight spec and the test suite it can check itself against.
Dex: So how do you put this into action.
Nova: This is the part to actually do. The lesson is not "fire your engineers." It is "the spec is the product now." Step one: pick one self-contained job you keep handing off, a data parser, a report generator, a migration script. Step two: before you let any agent touch it, write the spec as a set of example inputs and exact expected outputs, a checkable test suite, not a paragraph of wishes. Step three: hand the agent only the spec and the tests, no hints, and let it iterate against the tests until they pass. Step four: budget it like compute, not magic, watch the token spend, and cap it. The teams who win the next year are not the ones with the best model. They are the ones who can write a specification tight enough that the model can grade its own homework.
Nova: Your level up this week: install browse dot S H, point one agent at a website it keeps re-learning, and replace that re-discovery with a single installed skill. Ten minutes, and your next hundred runs get cheaper and more reliable.
Dex: If you want every story, every number, and every link from today in your inbox, the free Mainframe newsletter is the written version of this whole show. It is free, it lands every morning, subscribe and stay ahead.
Nova: That is today's Mainframe. The spec is the product. See you tomorrow.