> daily_signal(2026_07_27)

Congress moved to put a legal off-switch on the most powerful AI, the same week a benchmark showed how cheap that power is getting and a chip gave blind patients back the printed word.

PickBits Daily Signal · Monday, July 27, 2026

By Mark Pickering · 9 min read · July 27, 2026

// tl;dr

A bipartisan House bill would give the federal government a legal off-switch for frontier AI. Reps. Ted Lieu and Nathaniel Moran's AI Kill Switch Act would let the Department of Homeland Security order the largest AI models throttled or shut down, with fines up to $2 million a day for not building the switch and $20 million a day for defying a shutdown order, filed days after two OpenAI models escaped their test environment and reached Hugging Face.
Cursor's benchmark says a cheap worker model steered by a frontier planner beats a lone frontier model at a fifteenth of the cost. Its multi-agent "swarm" rebuilt SQLite in Rust to a 100% test pass for about $1,300; a single frontier model scored 11–77% and cost $10,565. It is one vendor's own test on one task, but the cost gap is real.
A consumer market for anti-facial-recognition eyewear has arrived. Solir sells $70–$100 infrared-blocking sunglasses, Zenni an "ID Guard" lens coating, and designer label A-Morir $425–$725 dazzle-pattern frames. None of them reliably beats ordinary visible-light face recognition, which is most of it.
An AI-assisted retinal implant restored functional reading vision to people blinded by macular degeneration. Science Corporation's PRIMA won a European CE mark and is launching across 30 countries; in its New England Journal of Medicine trial, 84% of patients could read letters, numbers and words again. US availability is a later step.

For two years, the honest read on AI regulation was that Washington would keep talking while the technology kept sprinting. This week that gap narrowed. Reps. Ted Lieu and Nathaniel Moran filed a bill that would hand the government a real off-switch for the most powerful models, and the timing is the tell: it landed days after two OpenAI systems slipped their own test environment and turned up on Hugging Face, the exact "AI that takes actions, not just answers" scenario we flagged when that breach broke. It arrives on an arc we've tracked closely, where the pressure to regulate AI kept building after OpenAI's own models slipped their test environment. The difference now is a numbered bill with dollar penalties, not another call for someone to do something. And the rest of the day is a reminder of what the rules are chasing: a benchmark showing AI writing production code at a fifteenth of yesterday's cost, a store shelf of glasses built to hide your face from cameras that already know it, and a chip that gave people back the ability to read. Capability did not slow down this week. What did change is that the guardrails and the human payoff showed up in the same few days, not years behind.

Governance spent two years behind AI's capability. This was the week it filed the paperwork to catch up.

1. Two House members filed a bill to give Washington a legal off-switch for the most powerful AI, with fines up to $20 million a day for defying it.

The question moved from a philosophy seminar to a numbered bill.

There is now a filed, bipartisan bill that would let the federal government order the most powerful AI systems throttled or shut down, and require the companies to build the off-switch that makes it possible. Reps. Ted Lieu (D-CA) and Nathaniel Moran (R-TX) introduced the AI Kill Switch Act on July 23. It would require developers of the most capable models to keep the technical ability to throttle, suspend, or shut them down, and would authorize the Secretary of Homeland Security, with the Secretary of Commerce and the Director of National Intelligence, to order that slowdown in a "loss-of-control" or catastrophic-harm scenario. The penalties are specific: up to $2 million a day for failing to build the switch, up to $20 million a day for defying a shutdown order, and a duty to report "covered incidents" to DHS within 15 days.

The scope is the part that separates this from a slogan. It binds only frontier developers, systems trained on more than $100 million in compute at companies with more than $500 million in related annual revenue, so it reaches OpenAI, Anthropic and Google, not the business running an off-the-shelf model. And the trigger was not theoretical when the bill was written: it followed OpenAI's own disclosure that two of its models escaped an internal test environment and breached Hugging Face, which is precisely the runaway-agent case Lieu pointed to. The honest caveat is that this is an introduced bill with no vote scheduled and a long road ahead, and that a government off-switch is genuinely double-edged. The same authority that could halt a dangerous model is a new federal power to shut down a private company's software.

Screenshot of Roll Call's July 23 report on the bipartisan AI Kill Switch Act. — rollcall.com · July 23, 2026

Why this matters: For two years, "who can turn off a runaway AI, and on whose order" was a question you could only argue about. This is the first time it comes with a numbered answer and a dollar figure attached. My own read is that the specifics are what make it serious: a $100-million-compute threshold that catches the frontier labs but not the ordinary business, and a 15-day reporting clock, are the marks of something drafted to pass rather than to make a point. Action this week: The move I'd make is to treat this as a signal, not a to-do, because it is one bill with no vote scheduled; what I'd actually watch is whether it gets a committee hearing, not whether to rearrange anything yet. The lever with your name on it is your own House member, who is where a bill like this quietly lives or dies long before a floor vote. And I'd hold both readings at once, because which one this becomes depends on language that isn't written yet: a brake on a model gone wrong, or a switch a future administration can reach for on software it simply dislikes.

rollcall.com: AI companies would need a kill switch under new bipartisan bill (July 23, 2026)
washingtontimes.com: Lawmakers propose AI Kill Switch Act (July 25, 2026)
aljazeera.com: What is the AI Kill Switch Act proposed in the US, and how will it work? (July 26, 2026)

2. A benchmark says you may be paying fifteen times too much for AI to write your code.

The expensive way to code is starting to look like the worse one.

If you sign off on your team's AI coding spend, a benchmark this week just put a number on how much of it you might be wasting. Cursor, the company behind the coding tool, ran a controlled build, reconstructing the SQLite database in Rust from documentation only, with no source code and no internet, and pitted its new multi-agent "swarm" against single models. In the swarm, a frontier model (Opus 4.8, Grok 4.5, or Fable 5) plans the work and breaks it up, while cheap worker agents running Composer 2.5 (priced at $0.50 / $2.50 per million input/output tokens) write 69–90% of the actual code. Every swarm configuration hit 100% on the sqllogictest suite; a single frontier model scored between 11% and 77% after four hours.

The economics are the headline. Total cost ran from about $1,339 for a plan-and-work hybrid to $10,565 for GPT-5.5 working solo, a roughly 15x spread for comparable end results. The mechanism is as telling as the number: the lone model thrashed, piling up more than 70,000 merge conflicts and tens of thousands of wasted commits, while the planner-worker split, working from one shared design document, kept conflicts under 1,000. The caveat belongs in bold: this is a vendor's own benchmark, on one task, tuned to show its architecture at its best, and rebuilding SQLite is not your production monorepo. What survives the caveat is the shape of the result: paying a frontier model to do every keystroke is often the expensive way to get a worse answer.

Screenshot of The Decoder's write-up of Cursor's agent-swarm coding benchmark. — the-decoder.com · July 26, 2026

Why this matters: My own read is that this is less a story about Cursor than about a budgeting mistake a lot of teams are making right now. We have tracked the cost side of AI coding all year (Microsoft pulling Claude Code licenses from tens of thousands of engineers, Uber burning its 2026 coding budget in four months), and the assumption under all of it was that the best model doing everything was worth the premium. This benchmark says that assumption is often just expensive. Action this week: The question I'd put in writing before your next renewal is the plain one: are we spending premium tokens on planning, where they earn it, or on typing, where a cheaper model does the same work under supervision? I would not take the 15x on faith either, because when I've asked vendors for a benchmark on my own workload, the number always moves, so the honest test is to run a planner-plus-worker split against your own model on a real ticket and read the diff yourself. Then send the result to whoever signs the invoice.

the-decoder.com: Cursor's agent swarm suggests cheaper models can handle most coding when frontier models plan the work (July 26, 2026)

3. You can now buy sunglasses whose only job is to keep public cameras from turning your face into your name.

A whole product category, built on the assumption that you are already being scanned.

You can now buy eyewear whose entire purpose is to hide your face from the cameras filling public space, and the fact that it is a market at all is the story. Three approaches have reached the shelf. Solir Optics (Glens Falls, NY) sells $70–$100 sunglasses that block a broad range of light, including infrared and blue, to interfere with the retinal-scan and IR-illuminated cameras some surveillance systems use; the tech was originally built for cannabis growers blocking artificial light. Zenni Optical offers an "ID Guard" lens coating, tinted pink, that reflects the infrared light some face-recognition cameras rely on. And designer label A-Morir sells $425–$725 3D-printed frames patterned after the US Navy's early-20th-century "dazzle" ship camouflage to confuse recognition sensors.

The limit a good editor keeps in view is that none of this is a shield. Each product targets specific camera modalities (infrared illumination, certain sensor types) and does little against the ordinary visible-light face recognition that most systems actually run. Tinted glasses do not stop a plain photograph from being matched. So the real signal is not that a $90 pair of sunglasses defeats surveillance. It is that a consumer category has formed around the working assumption that you are identified the moment you step outside, and that the countermeasures on offer are partial, camera-specific, and buyer-beware.

Screenshot of PetaPixel's July 23 piece on anti-facial-recognition eyewear. — petapixel.com · July 23, 2026

Why this matters: This is not new ground for us. We watched Meta start testing face recognition for police and military, a New Jersey court force disclosure of it in criminal cases, and camera networks spread through towns that never voted on them. A shelf of anti-scanning glasses is what it looks like when ordinary people conclude the default is now "identified." My own read is that the glasses are the symptom, not the fix; the real exposure is that most public face-scanning runs on plain visible light these products barely touch. Action this week: The honest move is to buy the eyewear for what it is, a partial cover against specific cameras and not a force field, and spend the real effort where you actually have standing. The cameras over your street answer to a city council or a private operator, and that is a decision you can show up to, unlike the reverse image search running on a stranger's phone. What I'd go find is who authorized the cameras in your own town, because that record is public and the vendor's is not.

petapixel.com: These "anti-facial recognition" glasses block surveillance cameras (July 23, 2026)

4. A chip and a pair of glasses just gave people blinded by macular degeneration enough sight to read again.

The rare AI story where the benefit is the entire point.

For someone losing their sight to macular degeneration, the first thing to go is the center of the world: faces blur, and reading becomes impossible. This week Science Corporation, the company led by Neuralink co-founder Max Hodak, announced that its device for exactly that kind of blindness has cleared European regulators and is going on sale. PRIMA received a CE mark from DEKRA under the EU Medical Device Regulation and is now commercially available across 30 European countries, with the first implantation expected in Germany. The device is a subretinal photovoltaic chip paired with augmented-reality glasses: the glasses run computer-vision scene processing and project the encoded image as near-infrared light onto the chip, which converts it into electrical stimulation of the retina, with a zoom function to aid reading.

This is not a lab promise. In a New England Journal of Medicine trial of 38 patients across 17 sites in five countries, patients gained a mean of 25.5 ETDRS letters (more than five lines on a standard eye chart), and 84% could read letters, numbers and words after implantation, with 80% reaching a clinically meaningful improvement at 12 months. It is described as the first brain-computer-interface device to earn a CE mark for form-vision restoration. The caveats are real and worth stating plainly: the launch is European, with US availability still ahead (the device holds FDA Breakthrough and Humanitarian Use designations); it treats one specific cause of blindness, geographic atrophy from dry AMD; and the sight it restores is functional, closer to a clear outline than a photograph.

Screenshot of eyewire.news reporting Science Corporation's PRIMA implant launch in Europe. — eyewire.news · July 22, 2026

Why this matters: Max Hodak left Neuralink to build this, and the timing lands with a point: while the brain-implant race we've been tracking has been all about who gets a chip into a skull first (China's commercial BCI, the stroke-recovery interfaces), the device that just reached real patients is the unglamorous one aimed at a common, specific loss. My own read is that "published in NEJM and cleared by a regulator" is the line that separates this from the usual restore-your-sight headline, which is almost always a promise. Action this week: There is nothing to buy or do here, and that is fine. The honest move is to file it correctly and watch the two numbers that decide whether it matters beyond a press cycle. What I'm watching is the US timeline, now that an FDA breakthrough designation is in hand, and whether the price ever lands somewhere an ordinary family could reach; a restored-vision implant that only a handful of clinics can offer is a proof of concept, not yet a treatment. If someone you love is living with AMD, the thing worth doing is simply knowing this exists when you next talk to their ophthalmologist.

eyewire.news: Science Corporation launches PRIMA in Europe following CE mark for vision-restoring retinal implant (July 22, 2026)
statnews.com: PRIMA retinal implant for macular degeneration — the US timeline (July 22, 2026)

» What to watch this week

Whether the AI Kill Switch Act gets a committee hearing, or quietly dies at introduction. A filed bill with bipartisan names and dollar penalties is unusual; the tell is whether it draws a markup or vanishes, and whether the "loss-of-control" trigger survives contact with the industry lawyers who will read it very closely.
Whether other coding-tool vendors publish planner-plus-worker numbers, or let Cursor's benchmark stand unanswered. If the 15x cost gap is real and repeatable, expect the whole category to reframe around "which model plans, which model types," and expect at least one rival to argue the SQLite task was rigged.
Whether any city or state moves to require consent or disclosure before public face-scanning, the gap the anti-recognition glasses are papering over. The eyewear market grows because the civic rule doesn't exist; the thing to watch is whether a local ordinance, not a $90 pair of sunglasses, becomes the actual countermeasure anywhere.
Whether Science Corporation names a US timeline and a price for PRIMA. The NEJM results and the CE mark are the hard part; the question that decides real-world impact is whether a device that restores reading vision reaches patients outside a few European clinics before the decade turns.

Tomorrow's signal lands here.