Every story, at length

Anthropic is warning that advances in small AI models could enable a new class of self-replicating malware targeting smart fridges, vacuum cleaners, and other IoT devices. The company says that by next year, smaller models may be able to autonomously self-replicate on such hardware, potentially giving rise to computer worms that propagate without human interaction and can infect devices they were never explicitly programmed for. The finding is tempered: Anthropic's own experiment revealed significant challenges in building such a replicator, and the company notes any resulting worm would likely be slow, with a doubling time measured in days.

The technical work centered on running small open-weight models on real or simulated low-cost, off-brand Android TV boxes and similar IoT hardware. Researchers used quantization to represent models with fewer bits and split workloads into shards distributed across devices. The Gemma4-26B-A4B model required just 15 boxes, each with only 4GB of memory, to achieve inference — meaning the model weights loaded and could generate output. The Qwen3.5-9B model was able to configure other LLMs to run on eight of nine hardware types tested.

What distinguishes this from traditional malware is adaptability. Conventional worms must be written for specific hardware or software targets. In Anthropic's tests, the model used its weights to make configurations on the fly for the specific devices it encountered. This suggests a replicator could figure out unfamiliar devices without prior programming.

Anthropic also warns that such a replicator would be difficult to shut down because there is no centralized point of control to disable. Unlike non-AI malware, it could adapt to new targets independently, though its propagation speed would likely remain slow.

HiRO-ACEv2 updates the two-stage High Resolution Output Ai2 Climate Emulator introduced earlier in 2026. ACE runs at 100 km resolution; HiRO sharpens those outputs to 3 km. The main change is HiROv2, a new downscaling model that predicts multiple high-resolution fields at once—rain, wind, and pressure—rather than rainfall alone, and shows improved skill on extreme events. When fed coarsened X-SHiELD data, HiROv2 matches the target simulation's high-resolution statistics to within 3%, even for events occurring once in a million data points. It also estimates total tropical cyclone energy accurately, including storms never seen during training, and generates hurricanes with realistic eyes and rain bands, though winds and pressure around the eye remain somewhat more lopsided than in X-SHiELD. Strong downslope winds are also reproduced realistically.

The updated ACE was validated over a 10-year simulation: hurricanes, average climate patterns, and distributions of rain, wind, and pressure all align with X-SHiELD. The full HiRO-ACEv2 pipeline then generated a full year of global 3 km data from ACE's own coarse simulation, not from real X-SHiELD inputs, and its statistics still closely match X-SHiELD. Across a 10-year run, hurricane energy is comparable to X-SHiELD over the same period.

Cost is the headline number: producing global high-resolution data with HiRO-ACEv2 costs roughly 300 times less than running X-SHiELD. If only one region is needed, savings can exceed that by more than a hundred times depending on area size. A streamlined HiROv2 is in development that could be more than ten times faster still.

Current limitations are explicit. HiRO learned from X-SHiELD, so it only works with the ACE version tuned to match X-SHiELD; it cannot yet downscale other climates such as a warmer future or ERA5 historical data. Planned work includes training on a 4°C-warmer X-SHiELD dataset and adding outputs like near-surface temperature and snow and ice. A web demo shows pre-generated HiROv1 and HiROv2 output side by side with X-SHiELD, and models plus tutorial notebooks are available for download.

The paper analyzes an agentic market where AI agents bid to claim each user prompt, with bids based on expected reward minus inference costs. First-price auctions are attractive for this setting because they are simple, transparent, and require no central knowledge of private capabilities or costs. But FPAs are not truthful: bidders shade their bids below true value, which can misallocate prompts and reduce welfare. The standard price-of-anarchy result for first-price auctions is a welfare guarantee of 1-1/e of optimal.

The authors identify two structural features likely to appear in agentic markets that push the bound higher. The first is a competitive floor: when multiple bidders usually value a prompt above some baseline share of optimal welfare—for example, because leading models are often distilled into cheaper near-equivalent alternatives—competition among them limits welfare loss, and the loss can be driven to zero as the floor rises. The second is near-independence of values after conditioning on the prompt, which holds when most predictable variation in expected value comes from prompt difficulty. As the value distribution approaches conditional independence, the efficiency guarantee approaches 1-1/e² of optimal welfare.

Both bounds improve on the well-known general-case result of 1-1/e. The authors validate the findings empirically using auctions with strategic bidders simulated on the EmbedLLM prompt routing dataset.

The contribution is theoretical rather than a shipped system: no new auction mechanism is proposed, and no production deployment is described. The novelty lies in showing that plausible properties of real AI markets—distilled model parity and prompt-driven value variation—materially change the worst-case efficiency picture.

Grace et al's 2024 survey of 1,580 AI researchers, mostly academics with no company affiliation, found over half put at least a 10% chance on AI causing "human extinction or similarly permanent and severe disempowerment," and 72% said AI safety research needed more priority. Geoffrey Hinton quit Google in 2023 to raise extinction-risk awareness, estimating a 10-50% chance; Yoshua Bengio said he "regretted his life's work." Former employees who left to warn include Paul Christiano (50/50 doom estimate), Daniel Kokotajlo (80%, forfeiting up to $2 million in equity by refusing a non-disparagement clause), and Jacob Coxon, who quit Anthropic in September.

The corporate-plot theory also requires implausible prescience. Dario Amodei debated superintelligence risk in a 62-page document dated October 27, 2013, before OpenAI or Anthropic existed and while he was outside tech. Sam Altman called superhuman machine intelligence "probably the greatest threat to the continued existence of humanity" in February 2015, ten months before starting OpenAI. Elon Musk warned "we are summoning the demon" in 2014, before his AI ventures, and donated $10 million after the first major AI safety conference. Shane Legg wrote his 2008 PhD thesis on machine superintelligence as "a nightmare scenario," a year before co-founding DeepMind.

If the slowdown proposal were a stock-pumping scheme, it failed. On the biggest cryptocurrency perpetual-futures platform, Anthropic's implied value dropped 4.57% and OpenAI's 13.42% when the pacing-the-frontier call was announced. Alex Tabarrok, head of the Center for Study of Public Choice, argues the facts don't fit regulatory capture theory: "Even a dull Ockham's razor cuts through that story to the obvious alternative: Amodei actually believes what he's saying." The distraction theory fares no better—long-term and near-term risk groups have formed a working coalition, jointly defeating Ted Cruz's bill to ban state AI regulation and co-funding the Tarbell Center for AI Journalism.

Alexander's conclusion: the warnings reflect genuine belief, not strategy. Many of the people issuing them made measurable personal sacrifices—forfeited equity, abandoned careers, donated millions—to get the message out.

The company announced Figma Specs at its first Config event in India: a Markdown editor where users define design language in a simple document that serves as a reference for both human teams and AI agents. Specs can exist as a standalone file or co-exist within the canvas, and users can reference other Figma elements such as canvas objects, embedded FigJam diagrams, code layers, and designs. Teams can also bring in design libraries, code libraries, and guidelines to manage context for the design system, which agents can consult to adhere to rules.

The move is explicitly framed as a response to the non-deterministic nature of AI. Figma made its agent on canvas available to a wider user base in the past week, and these new tools are meant to steer it toward more predictable outcomes. Separately, code layers — which add coding to designs on canvas for more interactive prototypes — move from private beta, announced in June, to public beta. Users can ask the agent to turn a design into code or change part of the code to iterate on a prototype.

Figma is also integrating Payload, the open-source headless CMS it acquired last year, with its Figma Make product for generating dynamic websites and directly managing content. The company said Payload downloads have increased sixfold in the last 12 months. Additionally, Figma is supporting over 20 new tools through its MCP, including Open Code, Dia, Grok Bot, and Indian AI startup Sarvam.

On the business side, Figma opened its first India office late last year with sales, partnership, go-to-market, and product support functions. The company said India is its biggest market by users outside the United States, and that it works with 40% of India's top 100 companies listed on the Bombay Stock Exchange.

Indie developer Popcar describes a wave of AI-powered decompilation of existing games, producing mashups between titles and a boom in preservation projects that get old games running on new hardware, often in a browser via WebAssembly. The piece, linked from Simon Willison's blog on 14th October 2026, frames the development as simultaneously a preservation win and an intellectual-property problem.

The preservation upside is concrete: old games can be ported, modified, enhanced, and made accessible on modern systems. Popcar notes this matters especially as console makers move fully to digital distribution and games become unplayable when servers shut down. Decompiled code gives users a path to keep running software they bought, independent of the vendor's continued support.

The downside is equally concrete. If AI can decompile games, the same technique applies to any software. Popcar argues the entire tech industry is about to change: competitors can extract and reuse code, paid applications can be open-sourced against their creators' will, and no one can stop it. The capability is not limited to old titles; it generalizes across the industry.

What is new here is not a single tool or benchmark, but the observed shift from isolated decompilation efforts to mass application. The article treats this as an emerging norm rather than a hypothetical, with preservation projects already shipping browser-based ports via WebAssembly.

The architecture separates four concerns: a golden dataset, an async inference runner, a judge layer, and a crash-safe logger. The golden dataset holds 20 to 50 cases drawn from real user logs or known failure modes, each with an id, an input prompt, and a short expected_concept string. The runner uses LiteLLM's acompletion with num_retries=3 and timeout=30, firing all cases concurrently via asyncio.gather and returning raw outputs only. The judge layer offers two modes: a deterministic string-match judge for fast, cheap checks, and an LLM-as-a-judge that uses Pydantic's JudgeScore model to enforce a typed passed boolean and reason field rather than free text.

The logger appends each scored result to a .jsonl file with a UTC timestamp, and a summarize function computes total, passed, and pass_rate. The deliberate separation of inference from scoring is the key design decision: raw_output is stored alongside the score, so refining a grading rubric later means re-scoring existing logs instead of re-running paid API calls. On a 500-case dataset against a paid API, the article notes, that saves real time and money.

Troubleshooting guidance covers three failure modes. LLM judge non-determinism is addressed by setting temperature=0 on judge calls. Sustained rate limits require a concurrency cap such as asyncio.Semaphore(10) around asyncio.gather. Judge model cost is controlled by running deterministic scoring first and routing only failures to the LLM judge.

Suggested extensions include adding latency and token cost fields to log records, wiring the harness into CI/CD to catch regressions on model updates, building an aggregation dashboard over the .jsonl files, and swapping in a rubric-based judge for pass/fail scoring. The article positions the harness as a foundation to build on rather than a black box to copy-paste.

Ricursive Intelligence founders Anna Goldie and Azalia Mirhoseini presented their approach at TechCrunch Disrupt on Wednesday, arguing that AI-driven chip design is a direct response to projected compute demand. Goldie cited an expectation of at least 10,000x more compute demand within five years, calling that figure potentially conservative. The startup's stated goal is not to scale data centers proportionally but to make chips, bandwidth, wafers, and fabs more efficient through automated design.

The founders previously developed AlphaChip, which produced a circular, donut-shaped arrangement of chip components that initially surprised engineers but proved successful. Ricursive is extending that direction by optimizing entire chips rather than partitioning designs into dozens of smaller blocks, an approach that yields unconventional layouts including wave-like patterns. The company says it plans to publish more details on those patterns in the future.

Goldie framed the shift as one where humans contribute ideas and creative direction while AI handles execution, a pattern she says is already occurring in software development. The economic argument is that if design time drops from six months to two years down to roughly a week, custom chips become financially viable for applications such as space, AR, VR, and biological computing, with energy efficiency as a further benefit.

The announcement is directional rather than a product release: no benchmarks, tape-outs, or customer deployments were disclosed. The novelty lies in the whole-chip optimization claim and the explicit framing of AI-designed silicon as a lever against data center build-out, but the underlying techniques trace back to the founders' prior AlphaChip work.

YouTube Music announced Wednesday a labeling system for tracks created with generative AI, covering composition and vocals. The label appears on the player page and, when clicked, shows the specific role AI played in the track's creation. The same labels extend to the main YouTube app. Partners declare AI use through one of three designations: Fully Gen AI, Partly Gen AI, or No Gen AI. Fully Gen AI covers tracks entirely generated by AI; Partly Gen AI covers works where artists incorporated AI-generated components alongside their own live elements. Pitch-correction tools such as Auto-Tune fall under No Gen AI. YouTube said it collaborated with music partners and deployed its own detection tools to identify AI-generated music.

YouTube's approach is disclosure-first rather than restrictive. AI involvement alone does not make a track ineligible for monetization, and the platform continues to experiment with AI-powered creation tools, including Dream Track, which lets select creators generate soundtracks featuring AI-generated versions of participating artists' voices.

The announcement lands amid diverging platform policies. Apple Music plans to introduce "Made With AI" labels as part of its AI Transparency Tags later this year. Spotify already has AI labels plus a music spam filter. TIDAL labels recordings it identifies as entirely AI-generated, lets listeners disable them, and excludes them from royalty attribution.

Deezer has taken one of the stricter stances. In July it reported receiving nearly 90,000 fully AI-generated tracks daily. The platform labels detected AI-generated music, excludes it from algorithmic recommendations and editorial playlists, withholds royalties from fraudulent streams, and is working on an option to let users filter out AI-generated music.

The Sydney startup, founded in 2024 from research at Macquarie University, deploys AI agents that answer scam calls, reply to shady messages and emails, follow phishing links, and join social media groups. The agents impersonate people of varying ages, genders, accents, and languages, and the company reports they have wasted 17 years of scammers' time across 2.5 million conversations. The goal is not just time-wasting: agents extract bank accounts, crypto wallets, phone numbers, malicious websites, and the brands being impersonated — details normally learned only after a victim has been defrauded.

Customers include Commonwealth Bank of Australia, TPG Telecom, utilities, telcos, and government agencies. TPG has diverted nearly 590,000 calls to Apate's agents; Commonwealth has seen a 76% drop in publicly reported scam losses; and an unnamed Australian bank used outbound agents to disrupt 22 scam campaigns. Customers get an insights portal with analytics on scams targeting their end users, employees, brands, and products, and can assign agents specific missions such as obtaining a bank account number or crypto-wallet address. Text agents search forums, investment groups, and social platforms for scam activity, then enter conversations under fabricated identities.

Apate uses both frontier models and its own fine-tuned models, trained on agent-scammer conversations and Kaafar's Macquarie research. The 20-person company raised $8.15 million in seed funding in August from Lobby Capital, OIF Ventures, Investible, Concept Ventures, and Baobab Ventures, with just under $10 million raised in total. It plans product development and international expansion, including a London office and a larger U.S. presence. The company pitched at Startup Battlefield 200 at TechCrunch Disrupt this week.

The context is a sharp rise in AI-enabled fraud: the FBI received more than 22,000 complaints of AI-related crimes in 2025, with about $893 million in losses, over $632 million of that from investment scams using AI-run conversations and deepfaked voices and appearances. Apate's approach inverts the industry norm of post-hoc blocking — telcos blocking numbers, banks watching transactions — by engaging scammers directly and extracting intelligence before harm occurs.

At TechCrunch Disrupt, Flock co-founder and CEO Garrett Langley defended the company's Gunshot Detection product against what he called misinformation, telling editor-in-chief Connie Loizos that the system now does 'just gunshots, plain and simple.' He acknowledged that Flock had prototyped broader capabilities after customers asked for alerts about abductions or people being harmed, but said the company pulled back after concerns were raised. 'There's only a couple customers that were prototyping it with us, and we said, OK, great, it's just going to do gunshots,' Langley said.

Flock's own marketing materials contradict that framing. The company's website says its audio systems can 'detect street takeovers early to stop chaos before it escalates' and 'identify severe crashes in real time so help arrives faster,' with alerts for 'car crashes, fireworks, and sideshows.' An FAQ section states that 'in addition to gunshots, Flock Gunshot and Audio Detection can detect fireworks, sideshows, and other disruptive community events,' and that its 'detection library is continually expanding to include emerging threats.' When TechCrunch asked Langley about the discrepancy, he repeated, 'It's just gunshots.'

Spokesperson Paris Lewbel offered a more qualified account: the technology is 'primarily used to detect gunfire but is also capable of detecting fireworks, sideshows, and potential car crashes,' and is 'not designed to monitor conversations or analyze speech and does not currently detect human screaming or distress.' That last point reflects an earlier retreat. When Flock announced Gunshot Detection last year, it said the system would listen for 'screaming'; after privacy groups objected to voice monitoring, the company changed the wording to 'sounds of human distress,' then said it removed distress detection from the customers trialing it.

Langley said Flock's microphone count is comparable to ShotSpotter, which Wired reported had about 25,000 microphones as of 2024. The remarks landed amid broader backlash: protesters gathered outside the Disrupt venue, Texas and Florida have moved to end relationships with Flock, and Amazon's Ring has ended its partnership with the company.