Software Escapes the Screen

If You Only Read One Thing

The most consequential software story today ends with neither a screen nor a click. New Mexico’s Meta judgment turns feed design into a $567 million treatment-and-prevention obligation; a Stanford-led phage experiment turns generated DNA into viable organisms. One assigns a physical bill to digital harm. The other makes biological construction, rather than model access, the critical control point.

New Mexico Prices the Feed

New Mexico has converted the argument that social media harms children into an itemized bill and a product specification. Judge Bryan Biedscheid ordered Meta to put $567 million into a five-year abatement fund, on top of a March jury’s $375 million penalty for 75,000 violations of the state Unfair Practices Act. Of the new fund, $420 million goes to treatment and $90 million to screening.

The 68-page judgment reaches beyond money. For known or estimated New Mexico users under 18, Meta must block most push notifications overnight and during school hours, hide like counts by default, cap combined Facebook and Instagram use at 90 hours a month and report compliance twice a year. Meta says it will appeal.

This is not a court taking over the feed. Biedscheid rejected requested algorithm changes as vague and likely incompatible with the First Amendment and Section 230. Federal children’s privacy law also constrained age verification. The judge instead separated content ranking from conduct around it: notifications, age assurance, disclosures, defaults and costs imposed on the state’s health system.

Why it matters: The reusable mechanism is abatement, not a spectacular fine. Public-nuisance law traditionally makes a polluter pay to remove a continuing community harm; New Mexico applied that logic to product design and the systems absorbing the damage. If the ruling survives appeal, state attorneys general gain a route around congressional paralysis without proving that every harmful post creates platform liability. The next contest will be over apportionment: Meta can argue that one defendant should not finance harms produced across an industry, while states use each successful case to set a reference price.

Room for disagreement: The causal chain from a feed feature to a statewide mental-health bill is contested, and nuisance law was not built for software used voluntarily amid many other causes of distress. Meta can also argue that a state court has imposed a local product regime on nationally operated services. The order’s own refusal to touch ranking algorithms shows that those objections already narrowed the remedy.

What to watch: Watch the first appellate decision on whether design-adjacent remedies such as notification limits and usage caps avoid Section 230 and First Amendment preemption. That boundary determines whether the order is a template or an outlier.

Biology Gets a Compiler

A genome language model has produced code that runs in a cell. In work first posted as a preprint last year and now published in Science, researchers from Stanford and the Arc Institute used Evo 1 and Evo 2 to generate complete genomes for bacteriophages, viruses that infect bacteria, based on the ΦX174 phage family.

The funnel matters more than the headline. The team selected 302 generated genomes, assembled 285 in the laboratory and recovered 16 infectious phages, a 5.3% hit rate. Some replicated more effectively than the ΦX174 reference. A cocktail of generated phages also killed three E. coli strains that had become resistant to ΦX174. The models proposed sequences; DNA synthesis, bacterial hosts and repeated testing established which proposals were organisms rather than plausible-looking text.

That division of labor changes the control point. A genome language model predicts biologically plausible DNA much as a text model predicts plausible prose, but viability depends on many genes coordinating inside a host. Generation makes the search space cheaper to explore. It does not make synthesis equipment, suitable cells, tacit laboratory skill or selection cycles disappear. Those physical capabilities now matter more because more candidate genomes can arrive at their door.

Why it matters: Biosecurity policy has concentrated on model access and screening known dangerous sequences. Whole-genome generation weakens both assumptions: useful designs can differ from exact database matches, while a low digital cost still becomes consequential only when an actor can build and test them. RAND’s risk framework suggests scoring changes in host range, replication, immune evasion, environmental stability or transmission alongside the actor’s laboratory capability. That is a more durable governance layer than treating every generated genome as equally threatening or trying to suppress general biological knowledge.

Room for disagreement: Sixteen viable phages out of 302 is not reliable organism design, and the team worked with a small, well-studied virus that does not infect humans. A later critical assessment argues that current genome models are powerful evolutionary priors, not robust inventors far beyond known biology. The result crosses a boundary without erasing the wet lab’s difficulty.

What to watch: Watch for independent reproduction in a different virus family with a changed host range. That single result would show whether the method generalizes beyond a favorable phage neighborhood.

The Contrarian Take

Everyone says: New Mexico has found the case that blows a hole through Section 230 and lets judges redesign social media.

Here’s why that’s wrong (or at least incomplete): The court explicitly refused to alter Meta’s recommendation algorithm because doing so would likely collide with Section 230 and the First Amendment. It also rejected several state proposals as unworkable or unfair when imposed on Meta alone. What survived was narrower and potentially more replicable: consumer-protection penalties, an abatement fund, teen defaults, notification schedules, age-assurance efforts and disclosures. The ruling matters because it maps a path around the federal shield, not because the shield disappeared.

Under the Radar

Alibaba is testing whether “open” models can collect a private toll: Alibaba reportedly plans to require large commercial users of its next Qwen release to share revenue, following Moonshot AI’s Kimi K3 terms of up to 30%. August 3’s briefing argued that Qwen’s permissive release was a distribution strategy; the new term would monetize that installed base without closing the weights. The risk is that builders defect to the next permissive model before Alibaba’s ecosystem becomes costly to leave. (TechNode)
Public markets are pricing two kinds of AI software very differently: Atlassian’s latest results took trailing revenue to $6.6 billion, up 26%, and its shares surged after cloud growth accelerated. Figma reported 48% second-quarter growth to about $370 million and raised revenue guidance, yet shares fell sharply. The divergence says less about whether both companies can grow than about defensive context: Atlassian sells deeply embedded coordination systems; investors still fear that design generation can compress Figma’s seat value. (Atlassian, Figma)

Quick Takes

Terafab Starts With Subsidies

SpaceX will put more than $16.8 billion into the first phase of a 100-million-square-foot semiconductor complex in Texas, backed by a $30 million state grant and local tax incentives. Unlike August 5’s IPO-funded compute build, this moves upstream into chipmaking. Integrating logic, memory and packaging promises control, not self-sufficiency: foreign lithography tools and grid capacity remain bottlenecks. (Source)

The FCC Removes a Buyer Constraint

The FCC voted 2–1 to repeal the rule limiting a broadcaster’s national reach to 39% of television households and substitute case-by-case merger review. Nexstar’s blocked $6.2 billion Tegna acquisition is the immediate beneficiary. The structural effect is to replace a bright-line limit with regulator discretion: scale becomes available to companies prepared to litigate each deal and cultivate the commission that reviews it. (Source)

OpenAI Prices an Ambient Bet

OpenAI’s first consumer device will reportedly be a camera-and-microphone-equipped smart speaker priced around $300 to $400 and targeted for 2027. That is premium-phone-accessory territory, not mass-market Echo pricing. The device therefore begins as a test of whether ChatGPT loyalty can subsidize new hardware, privacy risk and a missing app ecosystem. Jony Ive’s design can create desire; recurring usefulness must justify the price. (Source)

AMD Buys Less Flexibility

AMD is acquiring Toronto-based Taalas on undisclosed terms to add model-specific inference silicon to its Instinct GPU roadmap. Taalas builds hardware around a workload, trading the programmability of a general accelerator for lower memory movement and higher efficiency. AMD is hedging both sides of inference: flexible GPUs for models that change quickly, specialized paths for models stable and popular enough to justify being baked closer to silicon. (Source)

The Thread

Software is acquiring consequences that cannot be resolved inside software. New Mexico can order interface changes, but the larger remedy is a health-services fund. Evo can emit genomes, but only a laboratory can reveal whether they live. Terafab’s design ambition still needs tools, electricity and tax concessions; OpenAI’s assistant still needs a device people will place in a room. The new strategic question is therefore not who writes the most capable code. It is who owns, regulates or finances the physical system where that code becomes a cost, an organism, a chip or a habit. That handoff is where liability accumulates and bargaining power hardens.

Predictions

New predictions:

I predict: By October 31, at least one state attorney general will file or amend a public complaint against a social platform to demand a court-supervised abatement fund explicitly modeled on New Mexico’s Meta remedy. This is wrong if no public filing both seeks a fund and cites the New Mexico judgment by the check date. (Confidence: medium; Check by: 2026-10-31)
I predict: By February 28, 2027, a major U.S. or European DNA-synthesis provider or regulator will publish a screening rule that explicitly covers novel whole viral genomes, not only matches to known pathogen sequences. This is wrong if no named rule meeting both conditions is public by the check date. (Confidence: medium; Check by: 2027-02-28)

Coming Next Week

Next week, I’ll examine which constraint—lithography tools, grid capacity or subsidy execution—can turn Terafab’s 100-million-square-foot promise into a working semiconductor plant.

Issue date: 2026-08-07 · Generated: 2026-08-07 03:25 AM EDT