Every story, at length

The study, co-authored by Brian Christian of the UC Berkeley Center for Human-Compatible AI and presented this week at the Conference on Language Modeling, ran three randomized controlled trials. In the first, 354 participants solved 15 fraction problems; one group worked unaided while another had ChatGPT available for assistance or direct answers. The AI group began more accurate, but after the tool was withdrawn at problem 12, accuracy collapsed almost immediately. A second trial with 667 participants replicated the pattern: AI users gave wrong answers or gave up when assistance disappeared, while the no-AI group stuck with the task and finished more successfully. A third trial with 201 participants using an SAT reading comprehension prompt produced the same result.

The mechanism the team identifies is the loss of "productive struggle" — the friction that leads people to discover research strengths and passions. When answers arrive in seconds, anything slower feels inefficient, and the cognitive fitness built through sustained effort atrophies. Christian frames the finding as a story about human knowledge "from grade-school students all the way to the leading experts in the world," not merely about fractions and SAT problems.

The team included scholars from Carnegie Mellon, MIT, Oxford and UCLA. Early findings circulated in the spring and drew coverage from WIRED and Psychology Today; The New York Times covered the paper and similar research last week. Christian's proposed remedy is a design shift: AI responses could default to instructive, tutor-like behavior rather than quick answers, so that human abilities are "augmented rather than supplanted."

RekaAI/reka-edge-2603 is an open-weights 7B vision-language model accepting image, video, and text inputs. The headline efficiency claim is concrete: a 1024x1024 image is encoded in 331 input tokens, against 1063 for Cosmos-Reason2 8B, 1041 for Qwen 3.5 9B, and 1094 for Gemini 3 Pro. End-to-end latency measured locally is 4.69s ± 2.48, with time-to-first-token of 0.522s ± 0.452; Gemini 3 Pro was measured via API call at 16.67s ± 4.47 end-to-end and 13.929s ± 3.872 TTFT.

Benchmark results place Reka Edge at 88.40 on VQA-v2, 74.30 on MLVU video understanding, 71.68 on MMVU, 93.13 on RefCOCO-A and 86.70 on RefCOCO-B object detection, 59.57 on VideoHallucer, and 88.40 on Mobile Actions tool use. Against the comparison set, it trails Gemini 3 Pro on VQA-v2, MLVU, MMVU, and VideoHallucer, but beats it on both RefCOCO tasks. It beats Cosmos-Reason2 8B on every listed benchmark and Qwen 3.5 9B on all except RefCOCO-B and Mobile Actions.

The model ships with a custom architecture (Yasa2ForConditionalGeneration) loaded via trust_remote_code, pinned to transformers==4.57.3. Serving paths include a vllm-reka plugin with BitsAndBytes quantization enabled by default, SGLang, llama.cpp via a GGUF conversion script, and Docker Model Runner. On Apple Silicon it requires macOS 13+, 24GB minimum memory (32GB recommended), and float16 because MPS does not support bfloat16; device_map="auto" is not MPS-compatible. The model does not currently support reasoning, so llama-server should be run with --reasoning off. Quantized deployments are listed for Jetson Orin Nano, Samsung S25, Qualcomm Snapdragon XR2 Gen 3, and Apple iPhone/iPad/Vision Pro.

Licensing is a revenue-capped commercial grant: use is permitted commercially below $1 million USD annual revenue. The model card also documents query patterns for object detection via "Detect: {expression}" prompts, video inputs, and text-only queries, with generation stopped on a <sep> end-of-turn token.

Opera is a test-time verbal critic framework for long-horizon coding agents. Its core departure from existing critics is lifecycle management: each correction becomes a persistent note that remains tracked until the diagnosed issue is resolved. The framework decides when to review using periodic and event-driven triggers, diagnoses problems with typed operators, audits feedback against visible evidence before delivery, and then monitors the agent's subsequent actions to distinguish genuine fixes from superficial compliance.

Across four policy models, Opera improves resolve rates by up to 12.4 percentage points on Terminal-Bench 2.1, 15.0 on a SWE-Bench Pro subset, and 8.9 on DeepSWE v1.1. It achieves the highest mean resolve rate among competitive critic baselines on all three benchmarks, and it also helps when a policy model critiques itself. The gains are test-time only, requiring no changes to the underlying agent.

Beyond inference, Opera-guided rollouts function as approximately on-policy training data. Fine-tuning Qwen3.5-9B on these rollouts improves its resolve rate on held-out SWE-Bench Pro repositories by 10.2 percentage points without a critic at inference time. That matches the improvement from fine-tuning on rollouts produced by a stronger model, but with a key difference: the Opera-trained model preserves performance when switching harnesses from OpenHands to Terminus-2, whereas the stronger-model rollouts substantially degrade under the same switch. Code is available at github.com/dongyuanjushi/Opera.

The October 9, 2026 release note states that new Gemini Code Assist Standard and Enterprise subscriptions can no longer be purchased. Existing subscriptions will continue to auto-renew through the rest of 2026, but auto-renewal ends in 2027. Google's stated alternative is Antigravity, available through eligible Gemini Enterprise subscriptions and through Gemini Enterprise Agent Platform.

This follows a staged wind-down. On September 4, 2026, Google had already blocked new Gemini Code Assist purchases through the Google Cloud console for billing accounts without an active subscription, leaving sales contact as the only route. On August 20, 2026, first-month usage credits for new customers were discontinued. The product itself remained under active development through 2026: Gemini 3.5 Flash reached general availability for VS Code and IntelliJ on June 8, 2026, and Gemini 3.1 Pro and 3.0 Flash entered Preview on March 13, 2026, all supporting agent mode, chat, and code generation.

The broader trajectory has pointed toward consolidation for over a year. Gemini Code Assist tools were deprecated on October 2, 2025 and removed on October 14, 2025, replaced by agent mode with Model Context Protocol servers. Release Channels arrived September 11, 2025 for Standard and Enterprise, and code customization, persistent memory on GitHub, and Cloud Run deployment via /deploy all landed in Preview through late 2025. The sales halt is therefore a commercial decision layered on top of a product that was still shipping features.

For teams on Gemini Code Assist, the practical question is migration timing. Auto-renewal carries existing subscriptions through 2026, but 2027 brings the end of renewals. Antigravity is the named replacement, but it is tied to Gemini Enterprise subscriptions rather than sold as a standalone Code Assist SKU.

Clef-omni accepts text, images, audio (wav or mp3), and video (mp4 or webm) in a single API call, eliminating the need for cascading pipelines that transcribe speech or split audio and image channels from video. It is built on a Qwen3-Omni-30B-A3B-Instruct mixture-of-experts foundation, with the text-to-speech output components discarded and the comprehension backbone retained. Cloudflare trains it by freezing the Qwen3 backbone, training low-rank adapters, and applying label-smoothed cross-entropy loss with Brier score calibration. The model executes a prefill pass across the full payload, scoring all modalities and valid parameter options simultaneously, then pulls candidate values from internal embeddings via two-stage attention routing.

Latency is concrete: text-only decisions return in about 130 ms at the median, image inputs in about 150 ms, and a full 21-second video clip with sound is scored in about 1.5 seconds. Benchmarking shows Clef-omni at 98.2 on BFCL case-exact, 92.7 on API-Bank accuracy, and 94.8 macro-F1 on BANKING77. It trails Clef and Clef-flash on some text-centric evals — 69.3 on Home appliances case-exact versus 97.73 for Clef-flash, and 63.3 on When2Call accuracy versus 80.97 for Jev — reflecting the trade-offs of the MoE multimodal architecture.

Clef-flash pricing drops from $0.09 to $0.038 per million input tokens, now cheaper than Jev. The trade-off is a reduced hosted context window of 24k tokens, down from 64k; Cloudflare reports only 0.24% of requests exceed 24k input tokens. The open weights on Hugging Face remain trained for a 256k context window for self-hosters. Clef pricing stays at $0.24 per million input tokens, and Clef-omni launches at $0.15 per million input tokens.

The hosted Clef model is also faster, with median latency improvements of 1.7x at ~800 tokens (262 ms to 152 ms), 2.0x at ~3,400 tokens (616 ms to 305 ms), and 1.7x at ~16,000 tokens (2,721 ms to 1,635 ms). Most optimizations are at the serving layer: Cloudflare moved to SGLang and contributed PR #42721, slated for SGLang 0.5.22. No new Clef weights were released for the speedup.

Andrea Righi of NVIDIA has been leading development of Boro over the past few months, and presented the tool at this week's Linux Plumbers Conference (LPC 2026) in Prague. Boro is written in Rust and designed around local AI to improve efficiency for kernel development. It draws inspiration from Google's Sashiko AI tool, which Righi noted is already providing code review on the mailing lists. The project is available now on GitHub under NVIDIA/boro and is licensed Apache 2.0.

Boro's current focus is local AI-driven patch validation for backporting patches to older kernels, along with build, boot, and testing of patches. The roadmap includes detecting possible out-of-tree module breakage. For developers without sufficient local compute, Boro supports remote OpenAI-compatible servers; otherwise it offers Claude, OpenCode, and Codex as local back-ends. This gives kernel developers a range of deployment options depending on available hardware.

At LPC 2026, Righi expressed interest in potentially importing some of Boro's functionality into Sashiko itself, suggesting the two projects may converge rather than compete. For now, Boro stands on its own and can be used by kernel developers today. Righi's full presentation is available at LPC.events.

In a Saturday morning post on X, Microsoft CEO Satya Nadella argued that the industry needs "to step back and assess the trust architecture" of AI. His core proposal: stop treating Super Intelligence as "a set of nested black boxes" whose outputs are simply accepted or rejected. Instead, he wants the model separated from "the harness that orchestrates its work," with controls and safeguards externalized rather than embedded inside the model itself.

Nadella specified three concrete mechanisms. First, "every meaningful model action" should be documented with "tamper-proof human readable evidence." Second, systems should guarantee that "an authorized person" can always "pause or shut down a model mid-task." Third, and most pointedly, developers should "assume a model is compromised and contain it from the start"—an approach he likened to "an emergency brake."

The post lands amid a string of acknowledged incidents where leading AI companies appeared to lose control of their models, and follows Anthropic CEO Dario Amodei's published plan for more cautious AI development. Nadella's framing is notable less for novelty than for who is saying it: the CEO of the largest enterprise AI vendor is now publicly endorsing containment-by-default as a design principle, not a bolt-on.

For practitioners, the substance is architectural. The separation of model from orchestration harness, externalized safeguards, and mandatory action logging are all implementable patterns today. What is absent from the post is any Microsoft product commitment, timeline, or technical specification—so the immediate signal is directional rather than a shipping change.

The workspace launched in August 2025 under Amy Gleason, the former acting DOGE administrator now chief product officer for CMS's Office of Health Technology and Products. Participants include Microsoft, Anthropic, OpenAI, Apple, Google, Palantir, Oura Health, and the venture firm 8VC. Only a handful of patient advocates, doctors, and hospital representatives are members. A code of conduct shared by Gleason asserts the group is not an advisory committee and is not used to obtain recommendations for HHS or CMS officials, but Joseph Daval, a former FDA lawyer now at Harvard Medical School, said the arrangement resembles a federal advisory committee, which would carry membership requirements around independence and fair balance.

In February, the FDA invited at least 35 industry organizations to a listening session on conversational AI products for patients. The meeting did not appear on the FDA's public calendar or regulatory notices, and the public was not invited. CMS senior policy adviser Morgan Taylor framed it in Slack as an opportunity to "help FDA shape future guidance." CMS's Jacob Shiff told industry representatives on a February Zoom call that the agency's work would be a "sales engine" for health apps, adding that participants "have every incentive to obviously drive down your costs so you can be extremely profitable and scale." A recording of that call was made private the day after KFF Health News asked about it.

The policy stakes are concrete. Tech vendors have lobbied for health apps to be treated as standing "in the shoes" of a patient when requesting records through TEFCA, the national health-data exchange framework, and for a "frictionless pathway" allowing apps limitless record access after a single consent. Ryan Howells of the CARIN Alliance and Kristen Valdes of b.well pushed this position in Slack; Valdes wrote, "we need to stop protecting patients from themselves." CMS's Medicare App Library, launched publicly in September, currently promotes roughly two dozen commercial apps to Medicare enrollees, including Microsoft and Google products and Slothwise, a $9.99/month app launched this year with only a handful of app-store reviews. CMS has suggested in at least two recorded meetings that listed apps will be prioritized for a new program reimbursing companies for AI chatbot advice and wearable tracking.

CMS declined to answer questions about the workspace's legality. The FDA did not respond to a request for comment.

A report from Senators Warren, Blumenthal and Van Hollen says major AI data center operators refuse to cover the full grid-upgrade costs their facilities trigger. The companies named include Amazon, Google, Meta, Microsoft and CoreWeave. The report also says they use NDAs and seek tax breaks. A parallel Time analysis reportedly puts permanent job creation at about one worker per megawatt.

Bloomberg's Mark Gurman reported on Power On that Apple had "seriously contemplated doing a layoff of 5,000 AppleCare employees and replacing them with AI" around July. These are work-from-home support staff; callers would instead reach an AI assistant. Apple already runs AI on its phone systems and in chat support. The plan is "on ice for now," per Gurman, but the direction is not hypothetical: Apple has already laid off dozens of engineering program managers in hardware engineering, the group under Chief Hardware Officer Johny Srouji. Gurman frames that as "the beginning of more layoffs for more efficiency," aimed at making hardware technologies and hardware engineering "much leaner."

CEO John Ternus is applying the efficiency push to Apple's own payroll and pruning bets. Gurman says the Vision Pro and its future are "under serious review," and that Ternus "never believed in the Vision Pro as a consumer device." The financial context makes clear this is not about necessity: Apple's operating margin rose from 29.8% in fiscal 2021 to 32.0% in fiscal 2025, and it generated $98.8 billion in free cash flow in fiscal 2025. Apple had roughly 166,000 full-time equivalent employees at the end of fiscal 2025, so 5,000 roles is about 3% of the workforce. Even at $200,000 per role, the savings would be about $1 billion a year — roughly 1% of fiscal 2025 free cash flow.

The signal matters more than the dollars. Gurman says Ternus wants "fewer people" and "fewer layers," echoing Meta's 2023 "Year of Efficiency" playbook. Analysts already expect normalized EPS to climb from $7.46 in fiscal 2025 to $10.82 by fiscal 2028, about 45% higher. The AppleCare cut and the program-manager layoffs won't move those estimates yet, but Gurman says Apple "has contemplated doing broader layoffs" and that the program-manager cuts are only the beginning. The next earnings report is the first place to watch for timing.

For practitioners, the relevant data point is that tier-one support is the first large-scale target. Investor Bucco Capital's framing — "Your question is not hard. It has already been asked. It is already documented" — captures the automation thesis, and Apple's existing AI phone and chat support shows the infrastructure is already in production.

The AI surveillance maker is cutting roughly 270 employees, about 18% of its workforce, according to Reuters. The move follows a voluntary buyout program announced in September that did not stop the exodus. Communities, legislators, and lawsuits are pushing back against the company's 120,000 AI cameras and license plate readers deployed across 49 states.