Daily Roundup

AI border towers missed 1,050 deaths, MIT finds

Sep 22 roundup: MIT Tech Review documents failed virtual-wall surveillance; OpenAI forms math advisory; Meta Muse blocked on Amazon; EFF on EU Kids Act.

Top Stories

MIT Tech Review: 1,050+ deaths in range of border AI towers the system failed to catch

MIT Technology Review on Monday published a 15-month investigation called “Dying on Camera” finding that more than 1,050 people died within range of border surveillance towers between 2015 and early 2026, with the surveillance system repeatedly failing to detect them. The analysis cross-references nearly 4,000 remains-location records against about 600 CBP surveillance towers (catalogued by the Electronic Frontier Foundation out of roughly 800 the agency has deployed), according to the methodology piece. Three vendor systems are named: Anduril autonomous Sentry towers (more than 110 deaths near them since 2021), General Dynamics-operated RVSS camera towers (500 to 700 deaths since 2015), and Elbit’s Integrated Fixed Towers (nearly 350 deaths since 2015).

The piece documents broken equipment (66% of first-generation towers were unusable in 2024; 30% broken by January 2026), failed detection algorithms (groups misclassified as cattle; a Rio Grande car crash near a tower never registered), topographical blind spots (visibility as low as 10% to 25% of advertised area in some Otay Mountain Wilderness sectors), and unresponsive agents who sometimes “hang back” rather than intercept. Specific cases include José Morales Bernal, who died 360 feet from the closest tower in April 2024 from environmental exposure and was found by landfill workers, and the Lenda family - two-year-old Eloah and her father Daniel - who died in the Rio Grande near Del Rio, Texas in 2022 where three cameras had clear views. A companion policy-recommendations piece calls for a comprehensive death audit, better tracking of when tech leads to an apprehension, and treating any nearby death as a potential surveillance failure. The investigation was published in collaboration with Times of San Diego.

OpenAI forms a math advisory group after publishing a Navier-Stokes solution

OpenAI has formed an external math advisory group hosted at the Institute for Advanced Study in Princeton, the company said Monday. The group lands as OpenAI claims its AI has resolved more than 100 additional open problems across most areas of mathematics, and the company published a solution to the Navier-Stokes Millennium Prize problem. The advisory group’s role is intentionally narrow: a statement carried by TechCrunch reads, “the group will not be responsible for advising us on how to pace our internal progress on mathematics,” and a separate line clarifies, “Although we will give advice, we do not have decision making power at any AI company, and the responsibility for the decisions made by any company will rest with that company.”

The formation sits inside a broader, ongoing fight about how much credit AI deserves for math results. Critics have argued that benchmarks selected by frontier labs tend to favor pattern-matching over independent reasoning; the advisory-group framing is OpenAI’s attempt to put mathematician-shaped voices in front of that critique without ceding control of release pacing.

Amazon blocks Meta’s Muse agent from completing checkouts on Amazon.com

Users of Meta’s Muse AI agent trying to buy on Amazon.com have started hitting an error that reads, in part, “Continued access by an unauthorized AI agent violates Amazon’s Conditions of Use, to which our customers have agreed.” GeekWire first spotted the block. Amazon has not commented directly; the terms-of-use framing is doing the talking.

The piece spells out why Amazon has no legal obligation to host the agent: it operates its own foundation models and runs one of the most popular inference platforms on the internet (Bedrock), and if Muse places a bad order, Amazon would have to clean it up with both the customer and the vendor. Even Muse’s relatively low hallucination rate is “still pretty far from zero,” per TechCrunch. The Muse launch itself has been unusually fast by agent standards - Apptopia puts Muse iOS downloads in the U.S. and Canada at 1.8 million in the first 12 days versus ChatGPT’s 1.3 million, with U.S. mobile DAU at 642,000 versus 231,000 - but “outpacing ChatGPT” is the wrong frame for an agent that cannot actually buy anything on the largest U.S. e-commerce site.

Sherry Turkle: chatbots “deskilling” human empathy

MIT’s Sherry Turkle, in conversation with 404 Media about her upcoming book Artificial Intimacy: Who We Become When We Talk to Machines, argues that conversational companion chatbots risk being mistaken for genuine care, with the 404 Media writer framing them as “unfeeling, incapable robots” that users mistake for empathetic companions. Turkle’s verbatim framing in the conversation: “When people turn to a chatbot, they don’t have to dread the fatigue of genuine give-and-take” - a quote the interview uses to anchor her broader worry that avoiding genuine relationship work erodes the skills that make real relationships work.

Turkle is a licensed clinical psychologist and founding director of MIT’s Initiative on Technology and Self. The 404 Media interview does not name specific legislation, but the argument lands directly on the policy fault lines around AI companions - the EU Kids Act (covered below) and California’s teen-chatbot bills - both of which target the same use case Turkle is warning about. US states have already moved on the same question, with 78 active bills in 27 states following the Character.AI tragedies we catalogued earlier this year.

EFF: EU Kids Act would push services behind age gates, not protect kids

EFF argues in a Monday Deeplinks post that the EU Commission’s proposed Kids Act would replace one harm (children online) with a larger one (age-gated internet for everyone). The proposal phases access by age - under 13 no accounts, 13 to 15 restricted, 15 to 18 in a “safe-by-design” environment, full access for adults - and requires the EU age-verification scheme, including parental identity checks. EFF’s concrete worry: “age gates undermine civil liberties, reduce safety, and create barriers to internet entry, often at the expense of marginalized groups,” with no exemption for SMEs (only not-for-profit encyclopedias, scientific repositories, educational platforms, and open-source projects are carved out). EFF also notes the proposal skipped the typical impact assessment, and points to a French court ruling last month that struck down a similar blanket youth social-media ban as unconstitutional. Safety-by-design requirements cover social media, video-sharing, online games, AI companions, chatbots, and app stores - raising “fundamental rights concerns” EFF says are poorly suited to the decentralized Fediverse. The same logic that pushed Discord users off the platform after a facial-scan mandate is the political failure mode EFF is predicting at EU scale.

EFF files amicus brief in Levine v. FAA over drone flight restrictions

EFF is asking the D.C. Circuit to vacate an FAA temporary flight restriction on drones that effectively criminalized recording DHS officers (including ICE and CBP agents) and their vehicles - even from over half a mile away. EFF’s amicus brief in Levine v. FAA argues the restriction made observers and observers-and-journalists into potential felons for flying near federal agents. The ask lines up with EFF’s broader position that surveillance of the state is speech the First Amendment protects.

Quick Hits

  • Google ships the $899 Googlebook, a Gemini-native laptop. Magic Cursor for highlighting content and image work, Rambler for dictation cleanup, vibe-coded widgets, Gemini Spark and Gemini Live throughout. The piece frames it as a step up from traditional Chromebooks, with 12 months of Google AI Pro, 5TB of cloud storage, Gemini Advanced tools, and three months of YouTube Premium and Adobe Photoshop included. TechCrunch.
  • Tabby: a real-time bookkeeping interface from a former accountant, with 5,500 small-business users and about $100K ARR. Ahad Ali previously ran a 20-person firm doing 2,000+ returns a year; Tabby sits next to Sequoia-backed Rillet (which raised $100M in August) and Anthropic’s enterprise finance plugins. TechCrunch.
  • Ron Johnson, the architect of Apple’s retail stores, is skeptical of agent-led shopping. “AI will never be able to have you physically experience a product,” he told TechCrunch; agents make shoppers better-informed when they get to the store, not replacements for it. He calls himself an “AI optimist” but draws the line at AI owning the buying decision on big-ticket items. TechCrunch.
  • TypeSafe AI launches Jev, a “System One” / Decision Model LLM. Text in, typed probabilistic decisions out (yes/no statements, multiple-choice, numeric ratings). $0.042 per million input tokens with output free - cheaper than GPT-5 Nano ($0.05/million). Simon Willison.
  • Multiverse Computing reframes transformer block pruning as a constrained binary optimization (Ising) problem. Pairwise couplings from a Hessian over a second-order Taylor expansion; the technique reaches 76.9 MMLU vs. 54.0 for block influence at 50% depth on Llama-3.3-70B-Instruct, no retraining. Hugging Face blog.
  • Hugging Face open-sources relore, a repository-memory layer for coding agents. Indexes a repo’s history and connects past GitHub discussions - issues, PRs, review comments - to today’s code; excludes machine-generated content from the default evidence set; uses tree-sitter for Python symbol indexing. Apache-2.0. Hugging Face blog.
  • Hugging Face ships tokenizers v1 (release candidate). Three to 30x faster than v0.23 on a single Apple M4 Max thread across the benchmarks tested; 76% scaling efficiency across eight workers. Hugging Face blog.
  • Cloudflare Python Workers exit two-year preview and go generally available. Python runs compiled to WebAssembly via Pyodide in Cloudflare’s workerd runtime. Both multiprocessing and threading are non-functional in the WebAssembly VM. Simon Willison.

Worth Watching

  • Whether the MIT Tech Review investigation triggers a Congressional response. The piece lands with named failures, a four-point policy ask, and a timeline (2015 to early 2026). The test is whether the House Homeland Security Committee or the Senate Homeland Security Committee requests a CBP briefing within 30 days, and whether the Government Accountability Office reopens its prior “unreliable or insufficient” logging finding.
  • Whether the OpenAI math advisory group publishes any decisions before the end of 2026. The group has been framed as advisory only, not a release gate. The test is whether OpenAI publishes a math result with a dated advisory-group opinion attached, and whether mathematicians outside the group endorse or dispute that framing.
  • Whether other retailers follow Amazon’s lead on blocking Muse. Amazon has set the precedent that “violates our Conditions of Use” is sufficient reason to block an agent, and the platform also runs its own models. The test is whether Walmart, Target, or eBay publish blocking terms within 60 days, and whether Meta responds with a compliance roadmap or a counter-block of Amazon’s Bedrock-linked agents.
  • Whether the EU Kids Act draft picks up a small-business exemption or an impact assessment. EFF has called out both gaps; a French court struck down a similar blanket ban last month. The test is whether the European Parliament’s IMCO or LIBE committee files amendments covering either before the trilogue starts.
  • Whether the D.C. Circuit grants the petition in Levine v. FAA. EFF’s amicus brief is in. The test is whether the court schedules oral argument within the current term and whether the FAA withdraws or narrows the underlying flight restriction before that hearing.
  • Whether TypeSafe AI’s Jev pricing holds against the next OpenAI or Anthropic release. $0.042 per million input tokens with output free is unusually cheap. The test is whether a frontier lab publishes a Decision-Model-shaped product within 90 days and whether Jev lands a first named enterprise customer.