Saturday, 17 October, DeepSeek swapped new weights into the same 552B MoE, left the price untouched, and jumped Terminal-Bench 4.0 from 31.2 to 42.3. On Sunday, The Decoder reported that four Gemini 4 Argon coding agents had finished Minecraft end to end — 2,783 programs, 27 hours 14 minutes, about $520 in tokens, no human input. By Monday morning, an IMF staff paper had counted New Zealand’s exposure at six in ten workers, and a community college in rural Louisiana was graduating 24 fibre splicers a month toward Meta’s Hyperion build.
Four stories, one week, and they rhyme in a way none of them claims alone: the agent layer of AI — models, scores, token prices — is getting cheaper at a rate nobody would have priced in January, while the physical layer underneath it — power, fibre, cooling, the hands that cable a data centre — is hiring so hard that a four-week course has a waiting list. The money story and the jobs story have split. The agent economy is pricing everything down; the infrastructure economy feeding it is pricing labour up.
The Saturday shipping cadence is the tell
DeepSeek’s 1017 snapshot is the third Pareto move this year — September’s 75 percent flagship cut, the V4 open-weights release before it, now this: benchmarks up, task time down, price unchanged. Nothing about the numbers looks spectacular one line at a time. Together they move the cost of a completed task down twice: fewer tokens to finish, faster to finish, same price to run. That is the arithmetic the actual buyers feel, and it is why a “minor update” on a weekend is doing more damage to rival pricing tables than a new frontier model would.
The pattern repeats across the open-weights cohort this month. Meta shipped its first open weights on 15 October (Muse Spark Open, 816B/42B-active); Reflection’s Beam 1.0 weights went live on Hugging Face on 16 October under Apache 2.0. When we covered the September DeepSeek release, the point was that the “slow down” argument kept losing to a shipping schedule. Three weeks later the schedule is tighter: three major open-weight events in three days, each one resetting the floor price for agentic work and none of them waiting for a keynote.
What Minecraft actually proves, and what it doesn’t
The DeepMind Minecraft run is the week’s cleanest capability marker: four identical general-purpose coding agents, no game-specific training, no roles assigned, one paragraph of instructions — and the coordination primitives (task-splitting, shared memory, item recovery after a teammate dies, an Observer watching for spec-gaming and finding none) all worked. The deep pull from this morning’s source chain says something about where the hard part now lives. The Mineflayer library the agents drove — the “boring” community-maintained bot protocol underneath the frontier demo — sits at 7,601 GitHub stars and was pushed 11 October 2026, still under active maintenance by the volunteer PrismarineJS ecosystem. The agents wrote 22 percent of their programs into errors and still won. The frontier lab needed the plumbing; the plumbing didn’t need the frontier lab.
What the paper itself concedes matters just as much: structured state instead of pixels, Easy difficulty, a game the model knows from training, one success with no success rate. Minecraft’s goal is crisp, checkable, and cheap to fail at. Earlier benchmark work on Civilization VI showed the same class of agents stalling when goals turn vague, and the Princeton line of work showed them failing at open-ended research. The pattern holds: bounded, verifiable, sequential work is where agents are winning first — and that is precisely the work band employers are listing when they say “AI skills.”
The jobs story ran the other direction
Meanwhile the employment data refused the simple automation story all week. The IMF’s NZ mapping splits exposed roles about evenly between complementary and displacement-prone — only about one-third of workers in the vulnerable band, disproportionately female and mid-income. The Louisiana pipeline is blunt about which side is hiring: Meta’s Hyperion — nearly 10 million square feet — peak at 7,500-plus trade workers, about 1,000 permanent operational roles by 2036, a four-week course at roughly US$3,700 tuition and entry pay jumping from US$14 to as much as US$24 an hour. A job that requires no AI skills at all sits directly on the critical path of the AI economy: the concrete-and-copper layer is a hiring story even as the token layer is a price-cut story.
New Zealand’s edition of the split reads the same. SEEK NZ’s September report had job ads up 7.8 percent year-on-year with AI-skills mentions at a record 4.2 percent of listings — and the fastest-growing asks were governance and supervision, not prompting. The IMF flags NZ demand tilting toward applying AI: postings needing both developer and user skills have risen from about 24 percent in 2022 to about 34 percent in 2026. That’s the Minecraft lesson translated into job-ads language: the agents are general, but someone still has to own the objective, the environment, and the error rate.
Which side is reading the writing
The companies positioned on the right side of this week are legible from their actions, not their statements. DeepSeek keeps shipping cost-curve moves without a keynote. Google DeepMind spent its Sunday publishing the full telemetry of an agent run (the dataset is on Hugging Face, CC BY 4.0) because the argument it wants to win is that long-horizon agents are real and it holds the best evidence. Meta underwrote a US$5 million scholarship pool and a US$115 million Workforce Academy because, as one of its own policy directors put it, there is “a huge gap in the labour market when it comes to construction workers, power generation, grid modernization” — the company with the largest single data-centre build in history is spending to manufacture its own labour supply.
Who isn’t reading it: whoever is still pricing AI work as if the token curve were flat. Any business model that charges per attempt rather than per completed task gets repriced every time DeepSeek moves the floor — the same logic that made the September release an argument against a slowdown. Closer to the NZ ledger, the exposed-but-unprepared pocket the IMF identifies — mid-income clerical and administrative work, where UK official data put AI’s heaviest reported effects this month — is the band where “exposure” becomes real first. And the earlier think-piece thread about entry rungs thinning while the judgement layer prices up has now got a companion fact on the physical side: the rung being removed sits in the white-collar middle, while a parallel ladder is being built outdoors in Louisiana with a waiting list.
What this means in New Zealand
Practical reading, plainly. NZ has no gigawatt campus and, per the IMF’s own structural read — low R&D intensity, high electricity prices, shallow capital markets — is unlikely to build one soon. That makes the transferable asset the training architecture, not the building: employer-specified skills, weeks rather than years, a named job at the end. The IMF’s augmentation-side numbers say the strongest NZ upside sits in the upper-middle of the earnings distribution, which is another way of saying the governance-and-supervision skills SEEK sees at record demand are the ones worth paying to learn. And for developers, the cost curve means a thousand automated document-review tasks cost dollars in Western APIs and cents on the open snapshot — with weights on Hugging Face, nothing about that requires trusting a hosted provider at all.
🔍 THE BOTTOM LINE
- The agent economy and the infrastructure economy are moving in opposite price directions: task costs keep falling (DeepSeek 1017, Beam, Spark Open — three open-weight events in three days), while the physical layer underneath is bidding up trade labour (Hyperion: 7,500 peak construction workers).
- Capability is consolidating at the top of the stack (four general agents finished Minecraft; the 2011 bot library underneath, PrismarineJS/mineflayer, 7,601 stars, still volunteer-maintained) — the plumbing outlives any single frontier demo.
- NZ’s exposure picture (IMF: six in ten high-exposure, one-third vulnerable, women doubly exposed) and its hiring data (SEEK 4.2 percent record, governance asks fastest) both point to supervision-and-deployment skills as the defensible layer.
- The writing-on-the-wall test isn’t statements about AI, it’s budget lines: who’s funding training pipelines, who’s publishing cost curves, who’s still selling per-attempt.
❓ FAQ
Isn’t the Minecraft run just a stunt? Yes, partly — and that’s the useful part. The paper lists its own caveats (structured state instead of pixels, Easy difficulty, known game, single run, 22 percent of programs erroring). What it proves is that coordination across thousands of dependent steps is now within reach of general agents, at about US$520 in tokens. What it doesn’t prove is anything about open-ended work with vague goals — the other benchmarks from this year still show that gap.
Do cheaper agents mean fewer NZ tech jobs? The opposite direction, so far: NZ job ads are up 7.8 percent annually with AI-skills mentions at a record 4.2 percent, concentrated in governance and supervision. Lower task costs raise the number of tasks worth delegating, which raises demand for the people who deploy and check them — the band the IMF puts in the upper-middle of the earnings distribution.
Is the data-centre jobs boom relevant to NZ at all? The building isn’t — the IMF notes NZ’s constraints on capital and power. The architecture is: employer-specified, weeks-long, credential-backed training with a job attached. It replicates wherever employers can name the missing skills and fund the equipment, which NZ polytechs could run against grid and infrastructure projects.
Who benefits from this week’s open-weight moves in New Zealand? Anyone paying Western lab prices for high-volume agentic work. A snapshot at roughly one cent per million weighted input tokens — on open weights — makes the cheapest usable option something you can self-host, which removes both the vendor risk and most of the bill.
📰 Sources
- DeepSeek API Docs — V4.1-Flash-1017 release notes (17 October 2026)
- arXiv 2610.22319 — Gemini 4 Argon agents complete Minecraft (15 October 2026); The Decoder (18 October 2026)
- IMF Staff Country Report 2026/231 — New Zealand Selected Issues
- Business Insider — Louisiana Delta Community College data centre pipeline (18 October 2026)
- SEEK NZ Employment Report (September 2026)
- GitHub REST API — PrismarineJS/mineflayer repository activity (retrieved 19 October 2026)
- Hugging Face — DeepSeek model card; beat-the-game-minecraft dataset
