Every story, at length

The paper introduces MLExpert, a 35B Mixture-of-Experts agentic model for autonomous machine learning engineering, trained through a post-training method designed for long-horizon MLE optimization. The authors argue that feedback in MLE runs is costly and confounded: each experiment result reflects several preceding decisions, and a full run contains many such experiments. Their response is two-stage. Hierarchical agentic SFT transfers general tool-use and coding capabilities to MLE through outcome-guided supervision. Windowed Interaction RL (WIRL) then treats the multi-turn decisions between successive experiment results as an execution window and assigns it a shared outcome-based reward. Asynchronous collection and policy correction allow policy updates before a trajectory ends, addressing the credit-assignment problem across long interaction sequences.

To train the model, the authors built the Kaggle-MLE Factory, which constructs executable tasks and provides an interaction harness for model-directed experimentation. On MLE-Bench Lite, MLExpert improves the average medal rate of its base model from 37.9% to 72.7%. The paper states this is comparable to larger frontier models such as Kimi-K3 and GLM-5.3. Beyond the primary benchmark, MLExpert also improves performance on NatureBench Lite, agent scaffold optimization, and algorithm optimization, which the authors interpret as evidence that the learned optimization capability generalizes beyond the training distribution.

The contribution is less a new architecture than a training recipe for agentic MLE. The model size is modest relative to frontier systems, and the gains come from how supervision is selected from long trajectories and how intermediate outcomes are converted into rewards. WIRL is the central novelty: rather than rewarding a full run or individual actions, it rewards execution windows bounded by experiment results, which aligns reward granularity with the natural feedback cadence of ML experimentation.

FDE-Gym evaluates agents on end-to-end customer delivery rather than well-specified tickets. The benchmark comprises 95 delivery tasks drawn from 20 customer scenarios across 18 sectors, each adapted by industry experts from real-world delivery projects. Agents operate in simulated customer environments with scripted stakeholders and modifiable, production-like customer systems. Scoring is normalized between 0 (unchanged system) and 1 (reference solution), based on how much the agent improves the customer's own success metric, with penalties for stakeholder interactions including messages, trial runs, and waiting time. An online platform lets humans practice the same tasks under identical grading criteria.

Thirteen model configurations from seven providers were evaluated. In the realistic setting, where unstated constraints must be discovered without explicit guidance, the best configuration achieves a mean FDE-Gym score of 0.56 and resolves only 14 of the 95 tasks, where resolution requires a score of at least 0.8 with no incident. The authors identify discovering unstated constraints as the primary bottleneck: performance drops sharply when models must uncover them without explicit guidance.

Information discovery alone does not explain the gap. On the 23 tasks where the best configuration obtains every key fact from stakeholders, it reaches a normalized business uplift of 0.8 or above on only 10. This indicates that even with complete information, agents still fail to translate requirements into working solutions at a majority of tasks.

The benchmark's contribution is the shift from issue-report-style evaluation to delivery under ambiguity, live-system constraints, and stakeholder interaction. Existing benchmarks typically evaluate agents on well-specified tasks defined by issue reports, specifications, or task tickets; FDE-Gym measures something closer to the work of forward-deployed engineers.

The paper, submitted to arXiv on 15 October 2026, argues that modern MoE architectures combine fine-grained expert segmentation with shared experts, and that this pairing creates an unresolved tension. Fine-grained segmentation increases expressiveness by allowing more expert combinations at a fixed activation ratio, but it also enlarges the routing space combinatorially, which demands more exploratory routing. Shared experts — active for every token — reduce the variability of token representations, making routing less exploratory. The result is that the additional expressiveness offered by fine-grained segmentation may not be fully realized in practice.

The authors call this an "exploration dilemma" and propose MirrorMoE as a direct response. MirrorMoE shrinks the exploration space through token-adaptive suppression of unnecessary experts. Rather than letting the router search across the full combinatorial space while shared experts flatten the signal it needs to search effectively, MirrorMoE prunes candidates per token, reducing the space the router must explore and aligning exploration with the representational signal that remains.

The claim is empirical: MirrorMoE consistently improves performance across model scales. The abstract does not include specific benchmark numbers, parameter counts, or training budgets, so the magnitude of the improvement is not quantified in the source. What is new here is the diagnosis itself — the identification of shared experts as an exploration suppressant, and the framing of the problem as a conflict between two individually reasonable design choices. That framing is more specific than the usual observation that MoE routing is hard to train.

The proposed fix is architectural rather than a loss term or a routing-policy tweak. Token-adaptive suppression is a mechanism that operates at inference over the expert set, not a change to how experts are trained or how the router is initialized. That positions MirrorMoE as a drop-in modification to existing fine-grained MoE stacks, though the abstract does not specify implementation constraints, overhead, or compatibility requirements.

The provision, added in a recent update to Meta's commercial terms of use, stated the company could "use any content from [a business account's] domains, web pages, or URLs to provide, develop, improve, research, train, and maintain the integrity of" its products and services. Legal experts interviewed by Nieman Lab called it a "remarkable provision" that could give Meta permission to scrape the website of any publisher holding a Meta business account and use that content to train AI models.

Meta's public position was that the language consolidated existing crawling terms and applied only to ads and business products, not to how the company crawls the web generally. A spokesperson insisted the update did not reflect a change in crawling behavior. The Verge subsequently reported the provision has been removed, and Meta issued a statement: "We've heard you and want to clarify: the crawling terms we consolidated only applied to our ads and business products, and we've clarified that in the terms themselves. It does not change how Meta uses web crawlers."

The context is ongoing copyright litigation against AI companies. Meta defeated a high-profile lawsuit over alleged copyright infringement last year, but other cases against Meta and other AI firms continue. The removal appears to be a response to publisher and legal scrutiny rather than an admission that scraping practices changed.

A draft from Beijing's commerce and science ministries would add 40 technologies to the export-control list, 23 of them AI-related. The list includes large language models, agents, chip designs and computing clusters. Model weights and source code would be restricted above certain performance thresholds. API and SaaS access are exempted. Consultation runs to November 15. If passed, China's best weights would stop crossing its borders.

SpaceX announced a new Grok Voice model this week and disclosed that Starlink call centers are already using the technology to answer 20,000 customer calls a day. The company says roughly 70% of those calls are resolved without human intervention. Grok Voice agents verify customers, troubleshoot Starlink hardware, issue credits, and ship replacement equipment, and they can perform those tasks in multiple languages.

Starlink serves more than 6 million active customers worldwide, making it a substantial real-world deployment for AI voice agents. Rather than building a large human customer-service operation, SpaceX is betting that AI can serve its multilingual customer base directly. SpaceX did not respond to Business Insider's questions about which languages Grok supports or what categories of calls still require a human.

The announcement lands amid a broader rollout of production voice models. OpenAI released GPT-Live-1 with call-capable support in July, Google has launched similar enterprise systems, and Meta's Muse can handle phone calls for users admitted to its beta. What distinguishes the Starlink disclosure is scale and specificity: a named deployment with daily call volume and a resolution rate, rather than a model card or a limited beta.

For practitioners, the unresolved questions are the interesting ones. The 70% figure is self-reported and unverified, and SpaceX has not disclosed which languages are covered, what the escalation path looks like, or how failures are measured. A 30% human handoff rate is still a substantial operational load, and the economics depend heavily on what those handoffs cost.

The agency issued civil investigative demands exceeding 50 pages each, requesting documents on security incidents, safeguarding practices, privacy policies, and marketing claims. The probe focuses on consumer protection. The FTC also plans to compel top AI executives to testify and will send demands to third-party AI organizations later. OpenAI says it will cooperate.

Schwab CEO Rick Wurster told Bloomberg Television on Friday that the firm blocks unauthorised AI agents that try to move client money, including Meta's Muse. The core problem is credential sharing: clients hand passwords to an agent that then acts on their behalf. "We can't control it, the client can't control it, something could go wrong and who's liable?" Wurster said. That liability question remains unresolved in the US, where Congress is still debating AI agent accountability.

Wurster said he receives a weekly report on these blocks, and that the big agents are stopped in the tens and twenties each week. He framed that volume as evidence clients are not yet ready to hand their accounts to AI. Muse is Meta's personal AI agent, which can book, buy, and negotiate for users.

Instead of letting outside agents in, Schwab will stand up model context protocol (MCP) servers: secure links that let a client's chosen AI model connect to Schwab systems under rules Schwab sets. Through those links, outside AI tools will be able to check balances and do research, but not move money. Schwab is also steering clients toward Charley, its own AI assistant announced on 30 September.

JPMorgan Chase CEO Jamie Dimon took a similar position on Tuesday's earnings call, saying the bank would not let outside AI agents act inside its systems and warning of "rogue agents", per Cybernews. JPMorgan already lets some third-party tools view account data, but not send money or pay bills. The moves come as US and European regulators question banks about their use of AI agents. Notably, MCP links have had security problems of their own, with Google and JPMorgan among those that fixed the same flaw.

CoreWeave's unsecured debt due 2032 traded below 80 cents on the dollar, a level typically read as financial distress. The stock dropped 15.5% in a day. Bloomberg Intelligence now prices roughly 60% odds of default. As the most leveraged name of the AI buildout, CoreWeave's stumble could ripple across its debt-laden peers.

Anthropic published a report Friday on how its own workforce has changed over the past year, framing the data as a possible leading indicator of how AI reshapes hiring and labor demand. The sharpest signal is at the entry level: the proportion of junior-level employees in technical departments fell from 21% to 14% since early 2025, while junior technical hires dropped even more steeply, from 27% to 9%. Technical teams now represent 56% of the workforce, down from 61%, a shift Anthropic attributes partly to business functions scaling and partly to coding tools letting non-technical employees build software themselves.

The report also quantifies a change in what work consists of. Meeting time per employee is up 20% year over year, and up 16% for technical contributors, which Anthropic ties to the growing importance of human trust and accountability as AI output volume rises. In an April 2026 survey, 27% of employees said Claude made them more productive, while 20% said their job had shifted "from actually doing work to managing Claudes." One engineer compared the experience to babysitting: every Claude brings concerns demanding equal attention. The report describes human attention and judgment as a bottleneck, with Claude producing a "deluge" of text that must be read and organized.

On productivity, Anthropic reports more code per person, more sales per rep, and more contracts processed. It acknowledges that more code does not necessarily mean better code or business impact, but says it has found no indication that code quality is declining. The company presents these internal trends cautiously, noting that other organizations may never adopt AI as thoroughly as Anthropic, which one employee described as having "eternal pressure to move as fast as possible and use Claude for everything."

Michael Smith of North Carolina was sentenced to 18 months in prison after pleading guilty to one count of conspiracy to commit wire fraud. Between 2017 and 2024, Smith created thousands of fake accounts on streaming platforms using fake email addresses and fraudulently obtained debit cards, then ran software to make those accounts repeatedly stream his music. At peak, he operated as many as 10,000 bot accounts simultaneously. To avoid detection, he spread streams across thousands of songs, and from roughly 2018 he worked with the CEO of an AI music company and a music promoter to generate hundreds of thousands of AI songs, receiving thousands of new tracks each week for upload.

The scale is quantified in the DOJ's comparison: in April 2023, Taylor Swift's entire catalogue received 9.3 million streams on YouTube Music from family plan streams, while Smith's bot accounts used family plans to stream his AI-generated music 80.9 million times in the same month. Total fraudulent royalties exceeded $8 million.

What is new here is the criminal precedent, not the fraud technique. Bot-driven streaming manipulation predates generative AI; the AI element is the supply chain — an AI music company producing songs at a volume that made the scheme harder to detect and prosecute. The DOJ framed the case explicitly around AI-enabled fraud, and the RIAA and IFPI both used the sentencing to signal that enforcement, prevention, detection and disruption will be pursued together. The IFPI pointed to its Streaming Integrity Initiative as the industry's coordinated response.

For practitioners, the operational lesson is that distribution-side integrity checks are now a legal exposure question, not just a platform policy issue. Anyone building tooling that generates or distributes audio at scale should expect royalty accounting, bot detection and provenance to be scrutinised under criminal fraud statutes, not only terms of service.