Daily Brief Jul 9, 2026
A Northern California federal judge blocked three US government moves against Anthropic in March; the injunction is preliminary, not final
This issue was not emailed. An automated pre-send check did not pass, so subscribers did not receive it. The text is archived here unchanged.
At a glance
1. Judge Rita F. Lin's March 26 order paused the "supply chain risk" label on Anthropic, the President's directive that federal agencies stop using Claude, and the defense secretary's blacklist for defense contractors, according to NPR and The Hill; the court's own order hasn't been obtained.
2. swyx, the Latent Space podcaster who also runs the AI Engineer conference, predicted in 2023 that AI Engineer job listings, then about a tenth of ML Engineer listings, would overtake them by about 2028; swyx has a direct stake in that call.
3. In a large-scale training test on US supercomputers, AMD's weak-scaling efficiency fell from 88.3% at 8,192 compute units to 53.5% at 32,768, as relayed by Jack Clark's Import AI newsletter; that is one paper's figure, not a like-for-like comparison with NVIDIA.
This issue draws on SecondSource's own research brief for July 9, 2026. One current item, a March 26, 2026 court ruling, entered today after a targeted check; the rest come from a 2023 archive backfill and are each marked with their original dates. No qualifying new event surfaced in the past 24 to 48 hours, so the main line carries a single evidence update, and Named commentary runs a look back because no new named commentary came in this week. This issue uses 13 outside sources, each with a link to check. The full edition of this issue is the archive of record.
SecondSource isn't a news digest: each day it hunts the AI firehose for the insights that matter and the practitioner judgments worth tracking over time, and shows how every item was verified. The point is always which judgment got harder and who's been right, not what happened today.
Today's main line
1. [Evidence update] (ruling of March 26, 2026) A federal court, not a political reversal, halted the government's three-pronged pressure on Anthropic
Why it matters: State pressure on an AI lab met a constitutional check in court, though the order is preliminary and a parallel designation is unresolved.
Read the full item
On March 26, 2026, Rita F. Lin, a federal judge in the Northern District of California, paused three government actions at once. The first was the designation of Anthropic as a "supply chain risk". The second was the President's directive that all federal agencies stop using Claude. The third was the defense secretary's blacklist barring the defense industry from working with Anthropic. The court found that Anthropic could succeed on three grounds: retaliation for protected speech, due process, and the Administrative Procedure Act, the federal law that governs how agencies act. The order carried a seven-day buffer, so it took effect around April 2 (NPR, 2026-03-26; The Hill, 2026-03-26).
Back in March, only commentators' secondhand accounts were available for this item; today's targeted check filled in the full line. Anthropic's own statement of February 26 described the supply-chain label as "a label reserved for US adversaries, never before applied to an American company," and said: "these threats do not change our position: we cannot in good conscience accede to their request" (Anthropic statement, 2026-02-26).
Not proven yet: the court's order itself hasn't been obtained; its wording comes via NPR and The Hill
Verification: The injunction itself rests on two independent newsrooms, NPR and The Hill, reporting on the same day, plus the affected company's own statement. The court's original order hasn't been obtained. The phrase describing the case as classic retaliation for speech is the media's paraphrase of the court, and is treated as a paraphrase. A legal analysis also says that a parallel designation under the federal procurement-security law falls outside the injunction and is still undecided on appeal. That analysis is a single source with no citable link, so it is read for direction only.
Where it stands (as of this issue): According to a web check on July 9, 2026, the preliminary injunction is not final, and the parallel track has no final ruling either. A separate March report said the Pentagon's chief technology officer stated publicly that the ban would still be enforced, so resistance inside the executive branch is real. The export controls (imposed June 12, lifted June 30, as corrected in the previous issue) are a separate front; the two run in parallel and don't contradict each other.
Judgment update: SecondSource's long-term judgment that state leverage over AI labs cuts through their commercial hedging gets a mechanism correction. The original judgment held that the only real constraint would come from political norms, and it left out the judicial brake: the pressure was not withdrawn, it was blocked by a court on constitutional grounds. The parallel track is unresolved and enforcement meets resistance, so this does not mean Anthropic is in the clear.
Named commentary
(Look back, June 2023) swyx's prediction of the "AI Engineer" profession picked up a second, independent line of evidence today. In The Rise of the AI Engineer (2023-06-30), swyx, co-host of the Latent Space podcast, argued from supply and demand. By swyx's rough estimate in the same piece, about 5,000 LLM researchers worldwide can train models, against about 50 million software engineers. That gap, swyx argued, will force a new profession that doesn't train models but turns them into products. The piece came with a scorable prediction: ML Engineer job listings then outnumbered AI Engineer listings about ten to one, and that ratio would flip within five years, around 2028.
The reason for looking back today: the overnight archive backfill turned up a second, independent line. Raza Habib, co-founder of Humanloop, a company that sells tools for building LLM applications, said in an interview (2023-09-29) that the people buying LLM tools from Humanloop had changed. They were no longer ML-savvy research engineers but product-focused people who care about results rather than methods and are willing to pay. In Habib's words, "natural language written instructions have become part of your source code." With two independent lines now in hand, this call joins SecondSource's tracked judgments.
The discount: Both speakers sit in the same podcast circle, and swyx organizes the AI Engineer summit, a direct interest in the prediction. The scoring point is whether the listings ratio flips before 2028.
Model watch
No new paper made this issue; the item below comes from the 2023 archive backfill.
[Trend watch] (October 2023, archive backfill) "RAG is a hack, but a very good hack": the document where the retrieval path started. Jerry Liu, founder of LlamaIndex, said it himself in an interview (2023-10-05): "RAG is basically just, just a hack, but it turns out it's a very good hack." In November 2022 Liu built GPT Tree Index only to get around GPT-3's 4,000-token context limit. Retrieval became the industry default partly because fine-tuning was then expensive and immature. The live question hasn't closed: whether knowledge belongs in model weights or in external retrieval. Today's arguments over context engineering are a continuation of that 2022 question.
Chips & semiconductors
[Trend watch] "NVIDIA's moat is software, not silicon": asserted in 2023, measured on a supercomputer in 2025. The first-hand field account comes from George Hotz (geohot), who builds tinygrad, an open-source machine-learning framework, in an interview (2023-06-20). Hotz called AMD's RX 7900 XTX (US$999, 123 TFLOPS) the best compute per dollar on the market at the time, yet a card nobody used for machine learning. Running AMD's own demo apps crashed Hotz's whole server. Only after Hotz emailed AMD CEO Lisa Su did a fixed pre-release driver arrive. When Hotz reported a problem to NVIDIA instead: "The NVIDIA guy replied to me within an hour."
The measurement came two years later. In a large-scale training test on US supercomputers using AxoNN (March 2025, as relayed by Import AI 402), AMD's weak-scaling efficiency fell from 88.3% at 8,192 compute units to 53.5% at 32,768. AMD's math libraries and framework stability both lagged. Jack Clark's verdict in Import AI: for AMD to break the monopoly, its software has to mature first.
Verification: The two pieces of evidence are two years apart, of different kinds (first-hand experience versus academic measurement), and don't cite each other, yet both point to software inertia. Resonance across years is one of the strongest patterns of evidence. Discounts: Hotz sells the tinygrad framework and the hardware built around it, and the experience is self-reported. The AxoNN figures are one paper's results as relayed, and AMD was tested at a larger scale, not in a like-for-like comparison. The direction is solid; the exact figures deserve a discount.
Judgment update: For any challenger to NVIDIA, the software framework and community come before compute per dollar; cheap hardware doesn't rescue poor software. This 2023 judgment hasn't been overturned.
From the archive
Gullibility is the real bottleneck for agents: three people, 18 months, one conclusion, and the same analogy. Agents can use tools now, so the question is why they still can't be trusted to act fully on their own. Three front-line people who don't know one another said the same thing independently. Researcher Nathan Lambert (Interconnects, 2023-07) judged that wiring agents straight into outside systems was a dead-end opening move. Ethan Mollick, a Wharton professor, wrote in the week GPTs launched that "AIs are incredibly gullible" (One Useful Thing, 2023-11). The same mechanisms that make an agent useful, access to email, shopping and payments, are the attack surface itself. Engineer Simon Willison added the evidence (interview, 2025-01). As soon as Anthropic's Claude computer-use demo came out, someone handed Claude a web page telling it to download and run a file; Claude did, and the file was malware. In Willison's account, the dumbest possible trick worked.
All three reached for the self-driving car as an analogy. Self-driving cars could give road demos a decade ago; what has held back wide use is the cost of mistakes and the long tail of situations, and agents have the same structure. A 2026 paper sharpened the point: an audit of ten models (arXiv, 2026-06) measured the correlation between capability and resistance to adversarial deception at only r=+0.232 (p=0.520, statistically inconclusive). Stronger models don't automatically become harder to fool.
Verification: Three sources spanning 18 months and three vantage points, none citing another, plus one 2026 measurement: one of the most consistent topics on SecondSource's watchlist. Limits: the audit covers only 10 models, and its authors call it diagnostic. A demonstrable breakthrough in training against prompt injection would reopen the question.
What it means: Agent products divide on whether an error can be checked and rolled back cheaply. Where it can (coding, research assistants), gains in model capability pay off directly. Where it can't (automatic payments, decisions), the product stays stuck behind the gullibility bottleneck.
[Trend watch] (September 2023, archive backfill) A 2023 interview with LangChain's founder lays out how OpenAI absorbed the framework's features one by one. The interview and its editorial introduction (2023-09-06), an episode featuring LangChain co-founder Harrison Chase, set it down in writing. The chat API (March 2023) took over conversation management, function calling (June 2023) took over output parsing, and the fine-tuning API (August 2023) took over part of the retrieval tooling. The introduction puts it this way: "With each update, OpenAI and other frontier model labs realign the roadmaps of this nascent industry."
Verification: This is a single interview, but it describes facts that cut against LangChain in an episode built around LangChain's own founder, which carries more weight than self-promotion. Three years later LangChain is still around, having shifted toward observability and orchestration, so the mechanism holding doesn't mean frameworks must die. For any framework, orchestration or middleware business, the day-one due-diligence question is the same: which items on the feature list could disappear at a model maker's next launch event.
Product moves
No product news this issue.
— SecondSource · generated by our research system · 13 sources · Replies go straight to the editors as feedback
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.