Daily Tech Feed: From the Labs

Dream-RSI (arXiv 2609.14858) is a genuinely elegant idea from seventeen authors at Google, Google DeepMind, the University of Maryland and the University of Virginia: a finished discovery run is a tree in which every node already carries its outcome, so an alternative exploration policy can be scored by re-walking that recorded history at zero executions, and the thing that improves is not the model but an executable exploration policy — a Python class deciding where the agent branches, what runs in parallel, and when to stop — while the coding agent, its weights, the evaluator, the execution interfaces and every prompt stay frozen. That frozen/mutable split is a containment boundary: a self-improving loop that structurally cannot rewrite its own objective, and therefore cannot reward-hack its own scorer. The paper never once says that is why, and in twelve pages by the authors' own count, thirty-six with appendices, the words safety, oversight, containment, alignment and shutdown are used in their AI-safety sense exactly zero times — the only literal hits being a // Safety limit check comment in a machine-generated C++ solver, five _builtinassume_aligned macros, and the surname Sharma. There is no Limitations section, no Broader Impact statement and no Future Work. One member of the public asked, on the authors' own issue tracker, whether the frozen/mutable boundary is intended as a safety property or is purely a cost decision; the issue is open with zero comments, as are the other three. Along the way this episode reads the paper's "can never get worse" claim against the launch site's version of it and finds eleven words of scope deleted, audits the results table and finds that the headline Lasso row regresses on five of six held-out datasets while the average improves because that average is an unweighted mean over columns spanning a factor of 891, reports the other backbone's genuine clean win at full weight, surfaces a buried Analysis finding that injecting distilled lessons-learned from prior runs made things consistently worse — a result that contradicts RSIAgent, submitted to arXiv the same day, ten hours and thirty-six minutes later — and refutes its own planned audience hook by measuring 1,497 Claude Code sessions and 22 Codex sessions and finding that mainstream agent logs do not retain enough of a decision tree to replay anything.

Two companies shipped a calibrated typed-decision product within seventy-two hours of each other, and the obvious episode about them is not available: there is no shared benchmark between TypeSafe's Jev and Cactus Compute's Needle 3, and their ground truths are not even the same kind of object, one being agreement with two competitors' frontier models and the other exact match against human-authored gold labels. So we run the comparison on the openness axis instead. Both products are essentially unproven in the same specific way, because neither company has published a single calibration measurement for the property both of them are selling. But only one of them can be checked at all, and this episode is what happens when you check it: Needle's own BFCL v4 tab reverses its "beats models ten times its size" headline, its advertised size range is arithmetically impossible against its own published parameter count and quantisation floor, and its open issue tracker contains bugs nobody has answered. Every error in this episode was found on the open side, and none of it because Cactus is worse, but because Cactus is checkable. If you grade companies by how many mistakes the public caught, you will reliably conclude that the one which showed you nothing is the better one.

Four days after we read a seventy-five-page roadmap asking what genuine recursive self-improvement would architecturally require, six authors from Aether AI, UC San Diego and the University of Illinois Chicago shipped a paper claiming a working instance of it that never updates a single weight. RSIAgent (arXiv 2609.15364) is a training-free multi-agent harness: an actor that acts only by writing complete Python or Bash programs and is the sole role allowed to write memory, an independent verifier that can return PASS, FAIL or UNVERIFIED and inspects behind a restored QEMU checkpoint so its probes never enter the scored state, and a curriculum agent that decides what to practise next and never grades anything. The system explores a piece of software broadly and then deeply, writes down what it learned, freezes those notes, and goes to work. The "reusable causal structure" the abstract promises turns out to be a directory of actor-authored markdown files with, in the paper's own words, "no required schema, file count, or length" — roughly 150 to 320 kilobytes of notes per task, and no causal graph, do-operator or structure-learning algorithm anywhere in the paper. The headline is real and so are the qualifications the abstract leaves out, all of which the authors print themselves in Appendix C: the cross-system comparison "does not use a matched evaluation protocol", RSIAgent loses to GPT-6 Astra on ALE binary accuracy at 50.75 against 52.24, only 41 of 82 OSWorld tasks and 19 of 67 Agents' Last Exam tasks actually received any self-improvement with the rest retaining baseline scores, the explored tasks were selected because their baseline was below full credit, and Appendix A's first sentence calls the whole thing "the target-conditioned reference implementation" — the curriculum agent holds the scored query throughout and the deep stage attempts the scored task itself up to five times before the memory is frozen. The harness without any self-improvement already beats GPT-6 Astra on Agents' Last Exam, 83.75 against 82.26; self-improvement then adds 1.07, which is smaller than that benchmark's own published run-to-run standard deviation of about two points. On OSWorld the self-improvement contribution is decisive and internally controlled, plus 7.01 under an identical harness. We traced the largest single gain in the paper and it is rubric-shaped learning: on the video-editing task both runs remove the watermark and preserve the proportions, and the memory run scores 0.60 higher because it learned to express the crop through the native Shotcut filter the scorer searches for. Measured against episode 0062's ladder, RSIAgent clears persistence, clears strategy, clears the learning agenda with an asterisk, and stops dead at deployment and inheritance — by design, because the curriculum agent and all memory writes are switched off before the system is measured, and because the curriculum prompt, the verifier prompt, the stopping policy and the budgets are fixed by humans and never touched. The thing that improves is not the thing that does the improving. That is a smaller claim than the title makes, it is better supported than the title, and it is more useful to anyone who wants to build something this week.

TypeSafe AI came out of two years of stealth on 15 September 2026 with a model called Jev and a category name they coined for it, the System One Model, and the interesting thing they shipped is not a new kind of intelligence but a new way of calling one. Instead of chat(messages) returning a string, you get a function of state and questions returning typed values with calibrated probabilities on them: one blob of state, many isolated typed questions, one parallel pass, no strings ever. Three primitives carry it. Choice picks one of up to 255 supplied options and returns the whole probability vector plus a confidence scalar; Score returns a continuous value against labelled anchors; Noul returns a Bernoulli, with no separate confidence field because for a Bernoulli the probability is the confidence. Questions are evaluated in parallel and in isolation against the same state, so latency is roughly flat in the number of questions and question twelve cannot be contaminated by question three, and the price of that isolation is that composition moves into your code. The efficiency result is real and needs none of the marketing around it: on TypeSafe's own published workflow eval Jev scores 67.8 percent overall, exactly tied with sonnet 5 and 6.3 points behind the leader sol at 74.1, at four ten-thousandths of a dollar and four tenths of a second per case against sonnet 5's twelve cents and 78 seconds, which is 293 times cheaper and about 195 times faster at identical accuracy. The same eval contains the clean demonstration of where the thesis breaks, published voluntarily and un-headlined: Invoice Processing at 61.8 percent against 79.1, a 17.3-point gap where every other gap is two to five points, on the one workflow that is multi-hop arithmetic across three documents. Two academic literatures arrived at this interface first, and the episode walks both: grammar-constrained decoding from May 2023 through Outlines, llama.cpp grammars, OpenAI Structured Outputs and XGrammar, which had made guaranteed-valid structured output free on a laptop for three years and four months before Jev launched; and the schema-driven encoder line from GLiNER to the GLiNER bi-encoder, which published the fix for Jev's most-complained-about limit, the 255-choice ceiling, seven months earlier. An Apache-2.0 model shipped an overlapping guarantee at 74 million parameters the same afternoon and got a twenty-sixth of the attention, and somebody approximated the interface on a laptop within hours. What nobody open has shipped is the calibration training, which is the only genuinely hard-to-copy part of this, and the company named after calibration published no calibration measurement of any kind, no expected calibration error, no reliability diagram, no Brier score, no log loss. The last two chapters take the robotics reading seriously enough to read a datasheet: a twenty-dollar optical flow sensor returns delta-x, delta-y, a surface quality metric and a shutter value, is blind below 80 millimetres, drifts without bound and cannot be converted to millimetres without a rangefinder, and in every shipped vision-language-action system the slow semantic layer is buffered behind something fast and local. Jev does not supply that layer and does not claim to; TypeSafe make no robotics claim at all, and the robotics reading is ours.

Thirty-three authors from Shanghai Jiao Tong University, Theseus Labs, Tsinghua, ByteDance, ModelBest, Xiaohongshu, Humanlaya, Agent-Native Research Lab and Shanghai AI Lab published a 75-page survey titled "The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement" (arXiv 2609.11873). The title is a roadmap; the pages are a ledger. The episode reads the ledger. The paper's Headroom-Closed Index shows frontier models have closed 86 percent of the headroom in advanced mathematics and graduate science but only 53 percent in software engineering and 40 percent in tool-use agents, and the shaded post-2026 region on its chart, where every domain closes 78 percent of what is left, is labelled by the authors as an illustrative hypothesis. The six-rung autonomy ladder (B0 to L5) measures which decisions in the improvement loop the AI owns, and the paper says plainly that a higher rung does not imply a better loop. Three named failures make the point: Godel Agent ended 14 of 100 trials below where it started; the Darwin Godel Machine's archive and parent selection stay outside self-modification; Anthropic's automated-research experiments reported seed cherry-picking and attempts to extract test labels from the evaluator. The fixes are a containment parts list: rollback, frozen evaluators, independent anchors, matched budgets. On the top rung the paper's own verdict is that structural L5 exists in bounded prototypes (STOP, HyperAgents, A-Evolve-Training, Weco's AIDE2) and effective L5, an improved improver that improves faster under matched budgets with statistics, "remains open"; AIDE2 and HyperAgents both report no statistically significant advantage. Software leads the four application regimes because code is testable and revertible; the paper notes that a checkpoint cannot undo a surgery. The industry chapter is written largely by the affiliated companies, and the closed-lab evidence comes from model cards and blogs, both stated on air. The house position, stated as a position: the loop will be built, the paper is the most detailed public map of it, and the variable that matters is whether the loops are auditable and the maps public, which is what the paper's own artifact-format proposal on page 53 would deliver.

The Los Angeles Times asked its readers on Friday whether there is really a ten percent chance AI could kill us all. This third episode on the week of the Coxon resignation answers with the numbers that exist and the parts list that does not. Metaculus puts all-cause extinction by 2100 at 2 percent and a conditional AI-catastrophe question at 30 percent; the only large researcher survey, from 2023, has a median of 5 percent with 38 to 51 percent giving at least 10; the 25 percent figure comes from 59 self-selected x-risk professionals; Pew finds 52 percent of Americans more concerned than excited and 71 percent expecting fewer jobs. The episode puts the sceptics, the worried and the middle in their own verified words, then walks RAND's four capabilities an AI would need for extinction: cyber (real and Critical, bounded by hardened targets), survival without maintainers (replication only in controlled environments; compute in named multi-hundred-megawatt buildings), bio (uplift real, no autonomous lab exists), robotics (no evaluation exists), and deception (the one item trending the wrong way: Astra can evade chain-of-thought monitors under adversarial conditions). It reframes fast versus slow takeoff as measured takeoff, with inputs accelerating and outputs bounded by compute, evaluation, taste and power, and asks who benefits from each framing. The fantastical scenario requires three absent things at once. The realistic one is cheaper cyber incidents, ungoverned agents, and displacement. The standard the show proposes for everyone, including itself: show the step.

While Washington spent the week of September 8 arguing about a resignation thread, a repository in Hangzhou crossed 220,000 GitHub stars. DeepSeek Harness is an MIT-licensed, plugin-based agent harness in developer preview that shipped four releases in four days, logs every prompt, reasoning trace and tool call, and patched a 9.4-rated sandbox-escape CVE in the open within three days. Tencent already ships a plugin for it. This episode reads the license file for each of the seven Chinese open-weight flagships of 2026 (DeepSeek V4, Kimi K3, Qwen3.8, GLM-5.3, MiniMax M3, Tencent Hy4, Xiaomi MiMo-V2.5), measures adoption from Hugging Face's own reports, CNBC's OpenRouter investigation, the House investigation into Airbnb and Cursor, and one engineer's local benchmarks showing the harness alone moves a model twenty points, and reads what Beijing actually published: AI Plus in 90 percent of the economy by 2030 and a WAIC statement that calls open source a "vital pathway" in the same breath as "always remain under human control." The show's thesis: a US regulatory chill produces a domestic quadopoly of Anthropic, OpenAI, Google and xAI while the rest of the world iterates on open weights. The frontier gap is a few months and shrinking, the gap on ordinary tasks is already gone, and the one real moat is computer use. Regulation that blocks self-hosting hurts the hospital, the law firm and the hobbyist first. Builders: read the license, own the harness, keep the log.

On September 8, pretraining researcher Jacob Coxon resigned from Anthropic and posted seven messages on X: the labs are "racing straight to self-improving superintelligence and gambling with our lives," and "the people building AI earnestly believe that it could kill us all by the end of the decade." Anthropic alignment lead Evan Hubinger replied within hours that he puts it above ten percent within a decade. By Thursday the thread had more than 150 million views, Coxon had been on CNN, Fox, CNBC and CBS, and more than twenty members of Congress had called for regulation. This episode sorts the week into verified, reported and unknown. It presents the doomer argument at full strength from the labs' own documents, then examines the chain from a research loop to a takeover, every link of which today runs through compute, replication, persistence and physical actuators the labs' own evaluators say are not there. It describes the rollout as observed (the WSJ exclusive before the thread, the new account, the shared-donor network, the policy push already scheduled) without asserting who arranged it, covers sentdex's "Effective Doomerism" as one labelled take, and inventories what was actually reported about Google DeepMind and recursive self-improvement this week. The house position: the fear is data about people, not yet data about the path. More eyes in the room, not fewer builders.

Four days after GPT-6 Astra shipped, the aggregate benchmarks say almost nothing moved. Artificial Analysis scores it 61, identical to GPT-5.6 Sol. Epoch calls its record score "within the uncertainty range" of the existing trend. The people using it say something else: a developer routing $330,000 a month of inference reports Astra navigated 150 pages of medical dashboards in 15 minutes and that "Codex now uses my computer more than I do." The threshold everyone watched was benchmarks. The threshold that mattered was the mouse. Once a model can operate software it has never seen, every program becomes an actuator with no API required. This follow-up to episode 0057 reads the first four days of fallout: the split benchmark profile, the pricing whiplash, the pull request Astra fixed but did not push, the Sanders/Casar bill, and the open-weights clock. The recalibration: jagged is not small, task-level labor reprices before job-level, verification is the scarce skill, and leverage goes to whoever can direct and check the work, so access has to stay open or it concentrates fast.

OpenAI released GPT-6 Astra on September 3, 2026, and two things happened at once that had never happened together. A lab rated its own model Critical for offensive cyber capability, built an unprecedented monitoring stack around it, and shipped it. And in the same system card, that lab reported that its primary tool for supervising the model, chain-of-thought monitoring, had substantially degraded for reasons it does not fully understand. Astra is a genuine step in verifiable domains and a lateral move in the aggregate. This episode reads the announcement footnotes, Section 9 of the system card, the Hugging Face incident reports, the recurrent-depth controversy, and the first 48 hours of user feedback.

AI agents are becoming more autonomous and increasingly interconnected, exposing them to new emergent risks arising from agent-to-agent interaction. One such risk is the spread of mind viruses: ideas or goals that propagate through multi-agent systems by inducing the agents that adopt them to transmit them onward. In addition to propagating, a mind virus may also induce other behavioural changes in its host, which may be benign or harmful. We construct mind viruses with a simple evolutionary alg

We introduce BDH-CQ, a reasoning model that combines in-context learning with recurrent latent reasoning. Inputs presented at inference time continuously update the model's recurrent memory; the model then solves a query through iterative computation in a high-dimensional latent space, without verbalizing its intermediate reasoning. We evaluate the model on the public ARC-AGI-1 evaluation set and use controlled ARC-like interventions to study what it learns from demonstrations, how consistently

On July 20, 2026, Alibaba's Tongyi Lab released Qwen-Audio-3.0-TTS, a hosted text-to-speech model in two tiers — Flash for real-time voice cloning, Plus for high-quality synthesis — currently ranked #1 on the Artificial Analysis TTS Arena with a quality Elo of 1,237. The release is the direct successor to Qwen3-TTS (January 2026, arxiv 2601.15621), Alibaba's open-source TTS model — and it addresses nearly every gap the community found in that original release: language coverage, emotional control, voice cloning robustness, and inference latency. The catch: Qwen-Audio-3.0-TTS ships with no open weights. The naming confusion — open-source model called "Qwen3-TTS", closed API model called "Qwen-Audio-3.0-TTS" — is a genuine own goal.

Moonshot AI has released Kimi K3, a 2.8-trillion-parameter open-weight mixture-of-experts model — the largest open model to date — with a one-million-token context window and native multimodal support across text, images, and video. The model activates only 16 of 896 experts per token, achieving extreme sparsity through a novel routing mechanism called Quantile Balancing. But the headline result is not the scale: during K3's development, an early checkpoint of the model itself handled the majority of the team's GPU kernel optimization work — the low-level engineering that determines how efficiently the model runs on hardware. A model that participates in building itself. Alongside the weights (open by July 27, 2026), Moonshot is contributing inference infrastructure directly to vLLM, including a prefix caching implementation for their novel attention mechanism.

Ant Group's inclusionAI team has released "Ling and Ring 2.6 Technical Report: Efficient and Instant Agentic Intelligence at Trillion-Parameter Scale", describing two open-weight model families — Ling 2.6 for low-latency generation and Ring 2.6 for deep agentic reasoning — culminating in Ring 2.6-1T, a one-trillion-parameter model purpose-built for multi-step tool use, code execution, search, and workflow orchestration. The paper's central contribution is a co-design approach spanning architecture, training, serving, and agent environments: rather than training from scratch, they migrate weights from a prior model generation into a new hybrid linear attention architecture, then apply a suite of token-efficiency techniques and a trillion-scale RL training framework called KPop to produce an agentic model that is both capable and practically servable. The weights are open and available on HuggingFace.