Epoch AI, a group that forecasts AI progress, says intelligence-explosion forecasts skip how hard the promised technologies are to build; that critique is one Epoch column

Today's main line

1. [This week] (published July 7) Epoch AI says intelligence-explosion forecasts only do half the math

Epoch AI is a research group that builds quantitative estimates of AI progress, and its trend estimates are among the most cited in the field. A July 7 piece in its Gradient Updates column (original, 2026-07-07) takes on the intelligence-explosion debate: the question of whether AI will improve itself fast enough to become overwhelming within a few years. Epoch argues that the debate has worked hard on the first half, how capable AI gets. The second half barely gets quantified: even with very capable AI, how hard is it to build the specific technologies the forecasts promise, such as nanotech or Dyson swarms? The column puts the gap this way: "it's probably much harder to develop a nuclear weapon than an axe." Epoch's phrase for that gap is "the missing half of AI futurism debates."

Epoch proposes filling it in: pick one technology, state the assumptions about AI explicitly, then estimate the time and resources the technology needs. Epoch can name only one public analysis of this kind. Damon Binder's work on post-AGI energy production designs a minimal solar power system and finds it "could double on the scale of weeks." That figure reaches this issue only through Epoch's summary.

Not proven yet: the weekly-doubling figure reaches this issue only through Epoch's summary

Verification: A single source. The column is a named methodological framework, not an established fact, and Gradient Updates states that its pieces are the authors' views rather than Epoch's as a whole. The weekly-doubling figure is relayed from one study, so only its direction is usable.

Judgment update: Any "AI 2027"-style timeline now has one more completeness test: whether it accounts for the second half, or only extends the capability curve outward. Today almost all of them do the latter. That doesn't make the forecasts wrong. It means their error bars are much wider than they look.

2. [Trend watch] (published June 24) Epoch AI's read of 1,604 Chinese AI job postings shows China and the US hiring different kinds of people

Epoch AI scraped 1,604 AI job postings from Chinese labs and analyzed their structure in a June 24 piece (original, 2026-06-24). The sharpest finding is about experience. Chinese postings ask for 1.6 years of prior experience on average (3.4 years among postings that state a number), while comparable US lab postings ask for 5.5 years. Campus recruiting makes up nearly 20% of the open engineering roles in the Chinese data.

Put plainly, Chinese labs are building a pipeline that hires new graduates and trains them, while US labs compete for a stock of already senior people. These are two opposite ways of scaling. Location splits the same way. 93% of Chinese postings with a stated location sit in Beijing, Hangzhou or Shanghai, with Beijing alone at 63%; in the US, 85% concentrate in the San Francisco Bay Area.

Not proven yet: every figure, location and experience alike, comes from this one Epoch scrape

Verification: Epoch ran the scrape and published its method, including sample sizes and snapshot dates. The location and experience figures all come from this one analysis.

Judgment update: This is a structural difference, not a short-term swing in the hiring cycle. For US companies, it points to a lasting premium for senior talent. For China, the 20% campus share is a bet on a home-grown training system, which will show results only in three to five years. That makes it a call that can be checked later.

3. [Trend watch] (relayed June 15) Cognition offers to back its coding agent with up to US$10M in credits, a third way to close the trust gap that slows agent sales

Cognition is the AI coding startup behind Devin, an agent it pitches as able to take a ticket and write the code. According to the June 15 edition of the newsletter Exponential View (source, 2026-06-15), Cognition says it will fund up to US$10M in credits if its AI agents fail to deliver the engineering value enterprise customers paid for. The original terms haven't been checked.

The offer matters because the biggest obstacle to enterprise agent adoption is less whether teams can use the agents than whether they can trust the results. The industry already had two ways to close that gap. In the first, the vendor sends engineers on site to customize the agent until it works. In the second, the buyer absorbs the cost of checking: one measurement found that for every 7 hours of work an agent saves, someone spends another hour checking it. Cognition's offer is a third way. It moves the risk of a bad outcome from the buyer's books to the seller's. All three aim at the same gap: agents don't yet deliver value reliably enough for the product to sell itself.

Not proven yet: the refund terms haven't been checked against Cognition's own announcement

Verification: A single relayed source: Exponential View relaying Cognition's announcement. The details of the refund terms haven't been checked, so the offer shouldn't be cited as a risk-free commitment. It firms up if real payouts or repeat-purchase data come out, or if a second vendor offers the same kind of guarantee.

Judgment update: Whether other AI vendors match this kind of outcome guarantee, or explain why they won't, is a reading of how confident each one is in its own delivery.

4. [Trend watch] (relayed June 15) The share of firms using AI misleads; spending inside the adopters is wildly uneven

The same June 15 Exponential View edition relays a figure from econlab (source, 2026-06-15). The top 1% of US firms spend US$7,450 per employee on AI each month, roughly 650 times the typical firm's US$11.38 (7,450 ÷ 11.38 ≈ 650). Adoption therefore isn't a yes-or-no split. Even among firms that use AI, spending per employee follows a very steep power-law curve, with a handful of firms accounting for most of it.

Not proven yet: the 650x is relayed; econlab's original data hasn't been reviewed

Verification: A single relayed source, a third party quoting the data. Read the direction, not the decimals.

Judgment update: The share of firms that use AI says little on its own; how deeply they use it and how the spending is structured is where firms divide. A Chinese-language data point already on our watchlist points the same way, though it rests on a single relayed source: ByteDance's Doubao chatbot has 345 million monthly active users, a number that sounds huge, but most of them are students and older people using it to chat rather than for work. User counts have never been a proxy for revenue; the mix of uses is.

Named commentary

Nathan Lambert (September 2025, in his newsletter Interconnects; back in this issue after a re-check of the original): coding is the last broad arena where frontier models are still improving in ways ordinary users can see, and the same model performs very differently in different harnesses. Lambert is one of the most cited analysts of open models and a former researcher at Hugging Face and AI2. In a September 2025 essay (original, 2025-09), Lambert observes that one model scores very differently depending on the harness around it. Lambert's point is that much of the recent progress sits in this product layer rather than in the model itself.

Why it is back today: nine key statements in the essay were checked against the original, and all of them match. They include the large gap for one model across harnesses, OpenAI's Codex passing one million merged pull requests, and a capability ladder running from 2021 to 2027. The essay is the supply side of main-line item 3: Cognition puts money behind its agent because the value sits in this still-unsteady product layer, not in the model's raw capability. One discount applies. Lambert leans toward optimism about the open-model ecosystem, so the framework is usable but the optimism deserves a markdown.

Model watch

1. [Measurement method] (June 17) Epoch AI drafts an occupational task list for AI research, aimed at a streetlight problem in how AI progress is measured. Epoch's June 17 piece releases a first version of a task taxonomy (original, 2026-06-17). It breaks a frontier AI company's research workflow into six categories and more than sixty tasks, each rated from 0 to 5 on how far current AI automates it. The name borrows from O*NET, the US Labor Department's database of occupational tasks. The target is what Epoch calls streetlighting: "extrapolating whatever happens to be easiest to measure." A single-axis measure such as METR's length of tasks an AI can complete on its own is easy to track but misses how research work is actually structured. Read with main-line item 1, Epoch is doing one thing this month: turning "when does the intelligence explosion arrive" from an argument into an engineering question. Only Epoch is doing this so far. Whether it marks a shift in how the industry measures progress depends on a second organization adopting the same approach.

2. [Cross-checked] (July 2026) Four independent reads agree that the moat sits in the orchestration layer around the model, not the model itself. The orchestration layer is the coordination code that connects a model to tools, permissions and workflows. Latent Space's daily dispatch from the AI Engineer World's Fair conference (source, 2026-07) quotes Natalie Meurer of the customer-service AI company Sierra: "In practice, most customer-specific work takes place at the orchestration layer rather than in the models themselves." A June 23 post from the Claude Code team, matching Anthropic's official announcement, says Claude Code wrote 65% of the code for product teams. That 65% is Anthropic's own figure about its own product.

No major new paper this week. The academic material added to the archive over the past week was mostly a backfill of older papers on chip architecture, proteins and energy, with no direct link to this issue's main line. None was newly published this week, or an older paper suddenly back in focus, in a way that ties to the main line.

Chips & semiconductors

1. [Trend watch] (June 24) The same Chinese job-posting study yields a chip map: domestic chips are taking inference first, while the heaviest work, pre-training, still runs on Nvidia. From the 1,604 postings (same source, 2026-06-24), Epoch reads a structure. Production inference in China still commonly runs on Nvidia, using its CUDA and TensorRT-LLM software. At the same time, labs are hiring people with experience on Huawei's Ascend and Cambricon chips. At least one model, the 16-billion-parameter image model GLM-Image, was trained entirely on domestic chips. Epoch's best guess: "Chinese labs use domestic chips pretty frequently for inference, but rarely for pre-training large models." Inference, running a trained model to answer requests, is the lighter compute load; pre-training, building a model from scratch, is the heaviest. Under export controls, domestic substitution is therefore advancing from the light end up.

Verification: This is Epoch's inference from job postings, not a statement from any chipmaker. The inference point agrees with an existing line on our watchlist that Ascend chips are used for inference. As of this issue, this is one June reading of job postings, and domestic chips move fast; the real threshold to watch is when pre-training crosses over.

Product moves

No product news this issue.

From the archive

1. Bill Gurley: low margins in China's model layer are designed by policy, through a mechanism more intense than the one in electric vehicles. Bill Gurley, a partner at Benchmark and one of Silicon Valley's most senior tech investors, made the argument on the BG2 podcast in an episode recorded around August 2025 (episode). In Gurley's account, margins in China's model layer are squeezed not as a by-product of a price war but because the five-year plan writes open source in directly, and the provinces compete to carry it out. Several open models coexist, so no one worries about a monopoly. Gurley's analogy: in the fierce competition among electric-vehicle makers, a company can't take a rival's car apart to improve its own, but with models it can. Open weights let every lab build on the others' technical reports, which makes the mutual lift faster than in electric vehicles. Gurley's conclusion is that model-layer margins are structurally flattened, so the moat moves to distribution: Alibaba Cloud, ByteDance's apps, Xiaomi's devices. Nathan Lambert adds an inside view: Chinese labs save the 80% of research compute they would otherwise burn on repeated work by reading one another's detailed technical reports.

Verification: A single source, an older episode from August 2025, so it is a direction to read, not a settled finding. One academic boundary applies: there is a limit to how far models can improve one another. Research finds that some traits passed through distillation do not transfer between different models (arXiv, 2026-06). Gurley's spoken claim holds in direction but overstates how much can be copied.

This explains a split visible in main-line item 2. Chinese startups mostly focus on models and software, while platform giants such as Alibaba and ByteDance place broader bets on robots, in-car systems and wearables: when the model layer itself doesn't make money, leverage moves to the giants that hold the supply chain and distribution. The price pressure China exports may also put a ceiling on global frontier-model pricing.

Sources & accounting

The past 24 hours. The routine fetch brought in about 50 new pieces: 43 US public-company filings and 7 podcast transcripts, 6 of them read in full and 1 dropped because the whole episode had nothing new. Tracked accounts on X were quiet (@sama, @elonmusk, @bhorowitz and others), and every original post was read. Those 50 pieces are what actually came in over the past day. Separately, this week's one-time backfill of historical material added 2,101 academic papers, 2,083 blog posts, 974 newsletters, 851 company filings, 533 industry analyses, 288 podcasts and 202 X posts, among others. That backfill is for lookup only, is not presented as news and is not counted in the 24-hour figures.

The sources we track. The judgments in this newsletter rest on 529 named voices and sources on SecondSource's long-term watchlist: 305 people on X (Elon Musk, Andrej Karpathy, Sam Altman, Greg Brockman, Mark Zuckerberg, Demis Hassabis, Lip-Bu Tan, Jeff Dean, Noam Brown and others), 90 podcast voices (Satya Nadella, Sundar Pichai, Dario Amodei, Sam Altman, Demis Hassabis and others), 46 newsletters (Dylan Patel, Jack Clark, Nathan Lambert, Dwarkesh Patel, swyx, Ben Thompson, Nathan Benaich, Ethan Mollick, Andrew Ng and others), 48 personal blogs (Karpathy, Simon Willison, Chris Olah, Jason Wei, Boaz Barak and others), 48 paper authors (Noam Shazeer, John Schulman, Percy Liang, Tri Dao, Matei Zaharia and others), 70 institutional blogs (NVIDIA, Google Research, Hugging Face, SemiAnalysis, Interconnects, More Than Moore and others), 26 companies' earnings reports and calls (Jensen Huang, Lisa Su, C.C. Wei and others), 23 keynote speakers, 51 media interviews, 4 YouTube channels, 2 books, 2 courses and 1 government source. No sources were added today. This issue uses 8 outside sources in the body.

— SecondSource · generated by our research system · 8 sources · Got a view? Reply and tell us

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.