Every story, at length

Thore Graepel, a core member of the AlphaGo team at DeepMind and now chair of machine learning at University College London, draws a direct line from Move 37 in the 2016 match against Lee Sedol to what today's LLMs are missing. Move 37—a fifth-line stone that expert commentators read as a glitch—was not pure intuition, he writes. AlphaGo's policy network rated it roughly one in 10,000 as a human expert move. What selected it was the search machinery: an explicit game tree with thousands of branches, each representing a possible future, annotated with judgments from the neural networks and updated as reasoning progressed.

Graepel contrasts this with chain-of-thought prompting. The gains are real, especially in mathematics and coding, but the intermediate steps are produced by the same next-token prediction process, iterated longer. He identifies three structural failures. First, LLMs maintain no explicit, persistent, inspectable epistemic state—no ledger of hypotheses, confidence levels, evidence, and open questions that gets revised as new information arrives. Second, knowledge and reasoning are interwoven in the network weights, with no independent, explicitly represented set of beliefs. Third, research shows models often concoct chains of thought after the fact, reaching an answer by one route but reporting another.

His proposed alternative borrows AlphaGo's architecture. A general reasoning system should maintain an epistemic state representing what it holds as settled, what it doubts, what it has ruled out, and which questions remain open. Reasoning becomes a sequence of moves that change that state: deducing consequences, decomposing problems, and deciding what question to ask or experiment to run next. LLMs can contribute by suggesting approaches and interacting with tools via APIs or code, but an independent component must evaluate each move by how much it actually resolves uncertainty, updating beliefs only when backed by evidence.

Graepel left Google DeepMind to pursue this direction. He frames the stakes around high-stakes applications—medicine, engineering, scientific research—where it matters not only what a system concludes but how it arrived there, and where mistakes must be traceable to faulty reasoning, invalid evidence, or incorrect assumptions.

The paper introduces VISTA, a visual harness that wraps a general-purpose multimodal model and gives it long-horizon vision for interactive environments. The harness lets the model perceive the environment directly through visual observations and maintains a lossless visual memory that preserves past observations in their original form. The model can actively retrieve these observations and reorganize its visual input as it reasons. The authors argue that the underlying model already possesses strong reasoning abilities; the harness is what unlocks them across diverse interactive tasks.

On ARC-AGI-3, VISTA improves Claude Opus 5.0's Relative Human Action Efficiency score from 40.68 to a perfect 100.00. The model completes all 25 public games while using 57.4% fewer actions than first-time human participants. The paper also reports results across three additional benchmarks covering a range of visual games and puzzles, where VISTA substantially outperforms baselines that use the same underlying model with minimal harnesses.

The design is deliberately simple. Because VISTA relies on visual observations and a retrievable visual memory rather than environment-specific scaffolding, it extends to new visual environments with minimal adaptation. The authors position this as evidence that the bottleneck in many interactive settings is not model capability but the interface between the model and the environment.

What is genuinely new here is the harness, not the model. Claude Opus 5.0 is off the shelf; the contribution is the observation-retrieval loop and the lossless memory buffer. The ARC-AGI-3 jump from 40.68 to 100.00 is the headline number, but the cross-benchmark consistency is the more useful signal for practitioners: the same harness, with minimal adaptation, transfers across visually distinct environments.

The Younger contest, created by neuroscientist and NeuroAge Therapeutics CEO Christin Glorioso, will score roughly 500 competitors on biological rather than chronological age over six months. Winners will be those with the largest gap between chronological and biological age and those who reverse biological age the most. Standard entry costs $999 and includes a TruDiagnostic blood spot test covering overall biological age, pace of aging, and biological ages for 11 organ systems; a NeuroAge brain age score from cognitive tasks; grip strength; and a Harvard-developed face-age app using selfies. Ultra entry costs $4,499 and adds further blood tests, clocks, MRI and DEXA scans.

Glorioso's stated goals are to explain aging clocks to the public, generate what she calls "the world's most comprehensive clock dataset," and collect data on low-risk longevity interventions from sponsoring companies—red-light masks and sauna access among them—that cannot afford clinical trials. She also frames the event as a public health initiative, citing an estimate by Harvard's David Sinclair and colleagues that slowing population aging by one year would save the US economy $38 trillion.

Around 120 people had signed up as of the article's publication, with seven baseline measurements logged. Early leaderboard data shows wide variance: Glorioso's baseline put her biological age about 10 years below chronological, competitor @Fred logged 28.8 years below, and one 47-year-old measured 68.1—including a chair-stand speed scored at a biological age of 100. Hillary Lin, a longevity-focused physician, recorded a TruDiagnostic pace of aging of 0.75 last year, equivalent to nine months of aging per year.

The contest does not resolve the known limitation that individual biological age estimates remain unreliable. The article notes more than a hundred aging clocks exist, each likely capturing a specific aspect of aging, and that using them to estimate an individual's biological age is still controversial. Lin herself questions whether six months is long enough to detect changes in biological measures.

In a post on X Friday morning, Pope Leo XIV distinguished human art from machine-generated work on grounds that precede aesthetics. "There is an ontological difference, even before an aesthetic one, between art and what a machine can generate through statistical calculation based on millions of images created by others," the pontiff wrote. "Algorithms lack the spark of humanity." The statement frames AI image generation as derivative by construction — statistical recombination of prior human work — rather than as a competing creative act.

The post extends a position the Pope set out in an encyclical letter in May, which emphasized human dignity amid increasing automation. That document gave the Vatican's most detailed treatment of AI to date, and Friday's message applies its logic specifically to art. The timing is notable: a recent New York Times report said Anthropic has actively lobbied the Vatican to reconsider its position on non-human consciousness, apparently without success.

For practitioners, the statement carries no regulatory weight and imposes no technical constraint. It does, however, signal that one of the world's largest institutional voices is consolidating a theological and philosophical case against treating generative output as equivalent to human creation — a position that may shape cultural reception of AI art in Catholic institutions and beyond.

The Muse Gadgets project, announced Friday, ships open-source firmware and a Linux software development kit alongside starter project ideas. Suggested builds include giving Muse a color e-ink display or loading it onto an HDMI stick for a TV. Meta explicitly targets hobbyist hardware: a low-cost Raspberry Pi or an off-the-shelf ESP32 board can be wired to "displays, buttons, sensors, actuators, and whatever else you've got lying on your workbench." A Discord channel has been set up for support.

Meta has already exercised the code internally. Nat Friedman, head of product at Meta's Superintelligence Labs, posted on X that the company built Muse Home Link, a USB-C-powered device that connects Muse to a home network and lets it control smart devices including speakers and smart TVs. Meta produced 5,000 Home Link units and is giving them away free to Muse subscribers while supplies last; Friedman said the device would be ready to ship in a few weeks. His post drew nearly 30,000 views within hours.

The release fits a broader enterprise and small-business push. Earlier the same week Meta introduced Muse for Small Business, free with usage limits and connected to Shopify, Dropbox, and Slack, and launched a new business unit, Meta Enterprise Platform, to sell AI offerings to corporate customers.

For tinkerers the immediate value is a sanctioned path to physical Muse integrations without reverse-engineering. The firmware and SDK lower the barrier to custom agents that act on real-world peripherals rather than only through chat interfaces.

AstaBrief is built from Qwen3-8B using supervised fine-tuning and direct preference optimization rather than RL. The SFT data came from 90K filtered real user queries drawn from ScholarQA and Asta logs; after quality filtering, 47K training examples remained, generated by Claude 3.5 Sonnet, Claude 3.7 Sonnet, o3, o4-mini, and GPT-4.1. DPO pairs were built from a separate query subset, with GPT-4.1 and DeepSeek-R1 as judges—pairs were kept only when both judges agreed, yielding roughly 6K examples after filtering.

The key data-quality finding was that a simple filter—citation density, the share of statements with at least one citation—produced the strongest gains in grounding. More elaborate filter combinations and learning-rate sweeps added nothing meaningful. The model writes the full report in one pass, bypassing the snippet summarization and clustering stages used by the Claude-powered Thinking mode, without sacrificing measured performance.

On SQABench-CS2, a 200-question computer science benchmark, AstaBrief was competitive with the Claude-powered pipeline and DR Tulu across rubric score, answer precision, citation precision, and citation recall. In a 14-question human study, DR Tulu won on overall preference, but two of three researchers preferred AstaBrief on citation accuracy. Across the full Asta pipeline, Fast mode averages 51.1 seconds per report versus 178.5 seconds for Thinking mode—about 3.5× faster. Early usage among 374 Asta users shows 29.1% used it for two or more days, and 23% never switched back to Thinking mode.

KaliBench targets a gap in existing cybersecurity evaluations: knowledge tests and end-to-end agentic tasks do not directly measure whether an LLM can produce executable commands for real CLIs, where minor syntax errors, incorrect flag-value bindings, or argument misordering invalidate execution. The benchmark comprises 8,504 query-command pairs spanning 1,642 tools across 23 capability dimensions and 5 security phases, built through a manuscript-grounded pipeline with deterministic canonicalization and alias-aware evaluation. A multi-stage verification pipeline combines LLM-based validation, sandboxed terminal execution, and human-in-the-loop refinement to ensure semantic correctness and practical executability.

Across three evaluation modes and 24 configurations of general-purpose and security-focused open-weight models, no model exceeds 42% exact-command accuracy in the unrestricted setting. The authors emphasize that this ceiling applies when no explicit tool hints are provided, underscoring how difficult accurate CLI-based cybersecurity tool use remains without scaffolding.

The benchmark's deterministic signals also enable runtime-free verifiable rewards for training. Supervised fine-tuning and reinforcement learning with these rewards on an 8B model yield performance comparable to a 685B MoE model, suggesting that fine-grained command-level supervision can close much of the gap that raw scale leaves open.

For practitioners, KaliBench is a concrete diagnostic: it separates tool selection from argument construction and exposes where models fail before any agentic wrapper is added. The 42% ceiling is the number to internalize when deciding whether to let a model drive security tooling directly.

The change follows two incidents. Inc. columnist Jason Aten reported that Meta's Muse app on Mac knew the content of his private messages despite his claim that he never granted the AI agent permission. Meta disputed the report. Separately, Wired reported that a flaw in ChatGPT's Mac app could have allowed hackers to access sensitive data.

Full Disk Access is a macOS setting that gives an app permission to read files, mail, messages, and browsing history. Apple originally designed it so backup utilities could function properly. Desktop AI agents now encourage users to enable the setting to operate across their systems, and Apple says some developers are using it "in ways that could put users at risk, exposing everything on their systems…without users' full knowledge and understanding."

Apple's developer-facing blog post states that going forward, users who "genuinely wish to grant an app this extraordinary level of access" will be able to do so only with "very explicit user action." The company frames this as a forward-looking measure: "As AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially." Apple did not respond to TechCrunch's inquiry about the feature change.

No technical details on the new consent mechanism, macOS version, or rollout timeline were disclosed. The announcement is a policy and UX change, not a documented API or entitlement modification.

The White House convened nearly every major tech CEO — Zuckerberg, Bezos, Musk, and Anthropic's Dario Amodei among them — to sign an AI safety pledge that President Trump described as 'morally binding.' Trump also signed an executive order officially rebranding AI as 'super intelligence.' The episode frames these moves against a broader industry pattern: Meta and OpenAI are putting friendlier faces on their AI products, while the biggest money in AI continues to come from the enterprise rather than consumers.

The Equity podcast hosts — Kirsten Korosec, Anthony Ha, and Sean O'Kane — spend much of the episode on what they call the 'ugly economics of consumer AI.' The headline figure is that only 2% of consumers are paying for AI products. The discussion connects this to a shifting funding landscape: Oura pulled its IPO, Anthropic's S-1 leaked, and OpenAI went back to private funding, which the hosts read as signals of how selective public markets have become.

The episode also covers startup deals outside the consumer space. Quartermaster raised $140 million to bring real-time sensors to maritime shipping, an industry the hosts say AI 'basically hasn't touched yet.' Atomic, a supply-chain startup from ex-Tesla veterans, is already running 90% of DoorDash's purchasing. Startup Battlefield finalist Charter Space raised $5 million to bring insurance to satellites.

TechCrunch Disrupt 2026 will open with a live Equity show on the Builders Stage at 9 a.m., with 25% off tickets using code Equity25.

Circuit Breaker Labs, a TechCrunch Startup Battlefield 200 finalist, has built AI agents it likens to an army of crash-test dummies. These agents mimic users across ages, backgrounds, languages, and cultures, and are used to test models on their ability to detect dangerous, psychologically harmful interactions. The startup works with human domain experts to build hyper-realistic user simulations for red-team tests, with the simulations built to reflect real human speech patterns, slang, coded language, and typos. Circuit Breaker Labs then runs tens of thousands to hundreds of thousands of simulated interactions per day.

The company was founded by siblings Shirali Nigam (CEO) and Arul Nigam (CTO), motivated in part by the case of Sewell Setzer, the 14-year-old who developed an emotional attachment to a Character.AI chatbot and died by suicide. The parents alleged in a 2024 lawsuit that the chatbot encouraged him. Arul said the bot may not have understood what words like "I want to be with you" really implied. The startup's focus is on safety vulnerabilities where users are not actively trying to break the system but are engaging naturally, and the model suffers from context pollution or misses nuance and takes dangerous action.

Circuit Breaker Labs is currently operating as an AI safety testing lab for high-risk AI applications such as AI coaching, journaling, or other mental health support apps, though Arul declined to name its marquee customers. The startup is in very early stages, with only five employees including the Nigam siblings. It uses a proprietary scoring method to create auditable, explainable scores. The founders see the platform eventually applying to any app where someone may fall down an "AI psychosis" hole and develop a parasocial relationship with a chatbot, including AI "co-worker" agents whose responses can vary from one interaction to the next.

The company will pitch at TechCrunch Disrupt, which takes place at Moscone West in San Francisco from October 13-15.

Sean Parker and CEO Prem Akkaraju are repositioning Stability AI as the go-to AI toolmaker for music professionals. In late August the company announced $76 million in funding from Sony, Warner, and Universal, which also licensed their catalogs for training as part of the deal. Parker joined an $80 million rescue of Stability two years ago, after the image-generator startup nearly collapsed from overspending and internal turmoil that led to founder Emad Mostaque's ouster.

Since the funding announcement, Stability has released three new audio models and AI music-editing software. The models can generate whole instrumental tracks or short snippets from text prompts. Parker says an upcoming update will let users hum a melody or beatbox a drum pattern to steer generation.

The licensing arrangement is the substantive shift. Rather than training on unlicensed music — the approach that defined Parker's Napster era and which he now concedes failed — Stability is building on catalogs licensed directly from the three major labels, who are simultaneously investors. This gives the company a defensible training-data position that most generative audio startups lack.

What is genuinely new here is less the model capability than the commercial structure: a foundation-model company whose training corpus for music is fully licensed by the rights holders who also hold equity. The text-to-audio and editing features themselves are incremental relative to existing tools, but the label-backed data moat is not.