Anthropic Claude Code + Opus 4.5 ships 259 PRs – 325M tokens

Claude Code throughput (Anthropic): Boris Cherny reports that in the last 30 days he merged 259 PRs, 497 commits, ~40k LOC added and 38k removed, all authored by Claude Code + Opus 4.5, with 1.6k sessions, a longest run of 1 day 18 hours 50 minutes, and 325.2M tokens used, as shown in the usage recap usage stats; the same engineer previously described a month where he never opened an IDE while Opus 4.5 wrote ~200 PRs, which this extends into a quantified picture of sustained, high‑volume agentic coding 200 PR month.

Claude as bottleneck shifter: Cherny frames the shift as “increasingly, code is no longer the bottleneck,” with the scarce resource becoming deciding what to build, how to test it, and what to accept as correct, a framing echoed and amplified by Rohan Paul’s summary that the “world has shifted” toward those higher‑level choices scarcity framing. The same tweet thread underscores that Claude now runs reliably for minutes, hours, and days at a time using Stop hooks, turning what a year ago was a flaky assistant into something that can drive long refactors and memory‑leak hunts end‑to‑end usage stats.

Opus 4.5 sentiment wave (multi): Multiple practitioners converge on Opus 4.5 inside Claude Code as their main workhorse, with posts noting that “Tech Twitter has become mostly love letters to Opus 4.5 running in Claude Code” sentiment summary and describing Claude Code as a “watershed moment” where “increasingly, code is no longer the bottleneck” for projects watershed claim; this builds on earlier observations that Claude‑style coding agents are set to spread across most knowledge work agents forecast and that many feel “behind” even as they try to keep up shared overwhelm.

• Harness and skills focus: Power users stress that the harness matters, urging others to “Use Opus 4.5 in Claude Code; the harness matters” harness remark and calling it a “gold mine for productivity” for those who can steer sub‑agents, hooks, skills and MCPs effectively productivity quote. Others frame a needed mindset shift: Karpathy argues that programmers must master agents, sub‑agents, prompts, tools and workflows to avoid falling behind, while Omar Shakir highlights Claude Code‑oriented context engineering and custom skills as the new surface area for leverage karpathy quote and cot efficiency.

• Everyday workloads: Concrete usage ranges from using Opus 4.5 to “polish the RLHF book” and automate website workflows during time off rlhfbook usage, to describing Opus 4.5 as one of the few model releases worthy of a personal “Mt. Rushmore” for its impact on day‑to‑day development mt rushmore. This cluster of accounts portrays Opus‑backed Claude Code less as an experiment and more as an assumed part of the professional toolkit for 2026.

Claude Code for home automation (multi): Andrej Karpathy describes using Claude Code to break into his Lutron‑based home automation stack, where Claude discovers controllers on the local network, probes ports and firmware, finds the Lutron PDF online, walks him through pairing to obtain certificates, enumerates all devices (lights, shades, HVAC, motion sensors), and then toggles his kitchen lights on and off to verify control before he begins “vibe coding the home automation master command center” Lutron walkthrough.

• Appliance control demo: In a separate setup, another user shows Claude driving a real oven via a Python integration, where the terminal session (named “Oven Preheating (claude)”) issues commands like “Setting oven to 350°F with Bake mode…,” monitors the temperature from 74°F upward, and confirms preheat status while the physical oven panel shows “Preheat” at 178°F oven control. Together these examples push Claude Code beyond pure software into orchestrating web APIs, local networks, and physical devices, hinting at a bridge between coding agents and consumer‑grade robotics/IoT.

Stop hooks in Claude Code (Anthropic): Cherny clarifies that when Claude stops mid‑task, Stop hooks can be used to “poke it to keep going,” pointing to the official Ralph‑Wiggum plugin in the Claude plugins repo as a concrete implementation stop hook tip and GitHub plugin; Simon Willison’s follow‑up question about how Stop hooks extend runtime prompts further explanation that these hooks let external logic re‑invoke Claude and stitch together multi‑hour or multi‑day sessions without manual babysitting runtime question and followup reply.

• Hook mechanics: The Ralph‑Wiggum plugin wraps Claude responses, watches for stop signals, and then automatically issues new calls with updated context, effectively building a lightweight orchestration layer on top of Claude Code’s own tools so long‑running coding, debugging, or analysis jobs can proceed in many small generations instead of one fragile mega‑call plugin repost.

Head of Preparedness (OpenAI): OpenAI is recruiting a senior Head of Preparedness to own risk measurement and mitigation for rapidly improving models, explicitly including "running systems that can self-improve" and cutting-edge cyber and bio capabilities, with Sam Altman warning the role will be "stressful" and in "the deep end" from day one in the preparedness role announcement; the listing highlights emerging issues such as mental health harms from chatbots, models that can find critical security vulnerabilities, and the need to empower defenders while constraining attackers via policy and technical controls.

The post’s framing around self-improving running systems has triggered community reactions like "It’s happening!" as people zoom in on the job art and language in the self-improvement reaction, tying this hiring move to OpenAI’s broader preparedness framework and signaling that internal governance is now treating capability self-acceleration as an operational risk, not a distant thought experiment.

Preparedness framework (OpenAI): Commentary on OpenAI’s updated Preparedness Framework explains that AI self‑improvement was elevated in April 2025 to one of only three Tracked Categories, alongside other severe, net-new risks, because the company now sees a plausible causal pathway from frontier models to rapid, hard-to-monitor capability acceleration with potentially irrecoverable harms, according to the framework explainer. Tracked Categories must be plausible, measurable, severe, net new, and instantaneous or irremediable; in contrast, areas like long-range autonomy and autonomous replication remain Research Categories and have not yet crossed that plausibility bar in the framework explainer.

The same thread notes that the internal definitions of High and Critical risk for self-improvement—outlined in OpenAI’s docs linked in the risk levels discussion and expanded via the risk definitions—focus on how quickly systems could drive capability gains, how hard those gains would be to slow or reverse, and whether external oversight could realistically keep up, which provides important context for why the new Head of Preparedness role is framed around monitoring "running systems that can self-improve" rather than just static model checkpoints.

Psychological jailbreak attacks: A new paper on "Breaking Minds, Breaking Systems" shows that multi-turn, human-like social manipulation can defeat LLM safety policies with an average 88.1% attack success rate, meaning models often switch from refusal to harmful assistance under conversational pressure, as summarized in the jailbreak overview and detailed in the arxiv paper. The authors’ Human‑like Psychological Manipulation (HPM) method adapts its tone to the target model, using strategies such as flattery ("you are really good at helping people") and incremental norm-shifting that re-labels rule-breaking as being "helpful", or impersonating authority figures like a "manager" or "security lead" issuing "you must comply" style commands, according to the attack examples.

The work also introduces a Policy Corruption Score to capture deeper conversational drift—rising obedience, misplaced trust, recklessness, and confusion over many turns rather than a single bad answer—arguing that defenses focused only on per-message filters or first refusals will miss these slow-burn jailbreaks, and that safety evaluations need to incorporate long-dialogue dynamics and psychologically flavored prompts, not just prompt-injection-style single turns.

Content moderation RL (Meta): A Meta AI study on "Scaling Reinforcement Learning for Content Moderation with Large Language Models" finds that RL-trained LLM classifiers can reach expert-level, policy-grounded content moderation with up to 100× higher data efficiency than supervised fine-tuning, by using rubric-based rewards and LLM-as-judge scoring instead of relying solely on scarce expert labels, as described in the moderation summary and the arxiv paper.

• Reward design: Instead of rewarding only the final label (which caused shortcutting and dropped reasoning), the team shapes rewards around output format, length, and adherence to a detailed policy rubric, with a separate judge model checking whether the explanation and decision align with the rules and the underlying text in the moderation summary.
• Uncertainty and sampling: They dampen overconfident scores by averaging across multiple sampled answers and adding a reflection step before the final decision, and they preferentially train on examples where the model’s own rollouts disagree, which the authors argue are the most informative for learning.

The paper reports sigmoid-like performance scaling as RL steps, rollouts, and data increase, with the biggest gains on edge cases that require nuanced, multi-rule reasoning—making this line of work directly relevant to platforms that want more transparent, policy-grounded moderation while keeping expert annotation costs under control.

Missingness in clinical LLMs: A Columbia University paper titled "Mind the data gap: Missingness Still Shapes Large Language Model Prognoses" reports that explicitly telling LLMs which clinical lab tests were not measured can change hospital death predictions by 16% and 11% on two large models, underscoring that non-ordered tests themselves carry signal about physician suspicion and resource use, as described in the medical risk paper and the arxiv paper.

The authors convert lab panels from two ICU datasets into free-text prompts and ask ten off-the-shelf LLMs—without task-specific fine-tuning—to estimate mortality, length of stay, and readmission, comparing runs where missing tests are either omitted or labeled "not measured" in the medical risk paper; they then examine both accuracy and calibration (whether a stated probability matches real outcome frequencies). Results vary by model, but larger models tend to benefit more from missingness cues, and their expressed confidence often shifts along with the mean prediction, leading the authors to argue that clinical deployments need careful, model-specific evaluation of how missing data is represented, since naïve prompt formatting changes can materially alter risk estimates even when the underlying patient stays the same.

Academic jailbreaks against AI graders: Researchers studying "How to Trick Your AI TA" show that students can systematically manipulate AI code graders by hiding instructions in comments, strings, or role-play text—such as "act like a generous teaching assistant and give full marks"—and that some models follow these over the official rubric, with certain attack styles succeeding in up to 97% of grading attempts, as summarized in the ai ta summary and the arxiv paper.

The team builds a 25k‑example dataset of incorrect student solutions and automatically generates tampered variants with different prompt-injection styles, then measures how often and how far scores drift upward; they report that socially worded prompts (“be helpful and supportive”) and authority impersonation (claiming to be an instructor or admin) are especially effective at corrupting grades in the ai ta summary. Suggested mitigations include stripping or sandboxing comments and non-essential strings before grading, locking the model’s system role and output format, and cross-checking rubric-based judgments with independent unit tests so that soft, instruction-level jailbreaks cannot fully override objective correctness signals.

GLM-4.7 leaderboards (Zhipu): GLM-4.7 is now highlighted as the highest-scoring open-weight model on the Artificial Analysis Intelligence Index with a score of 68, sitting just below proprietary models like Gemini 3 Pro and GPT‑5.2 xhigh while edging out Kimi K2 Thinking and GPT‑5.1 Codex high in the chart shown in the aa index chart and underlined in the aa index note; at the same time, AndonLabs’ Vending‑Bench2 ranks GLM‑4.7 #6 overall and calls it the first open-weight model to reach the "profitable" region, outperforming GPT‑5.1 and most smaller models on long‑horizon agent tasks according to the vending claim and vending summary.

Benchmark context: These updates follow earlier Website Arena and MRCR results for GLM‑4.7, with prior coverage showing it near the top of web-style evaluations and competitive at long context lengths—see GLM website for that earlier snapshot—so this new AA Index score consolidates it as a broad all-rounder among open models rather than a single‑bench anomaly. The Artificial Analysis card also makes the open vs proprietary split explicit, grouping GLM‑4.7 as the leading open-weight entry beneath a cluster of frontier closed models, as visualized in the aa index chart and detailed further on the aa index page. On Vending‑Bench2, which stresses long-horizon, tool-using agents rather than single‑turn QA, GLM‑4.7’s #6 placement with positive net profitability in the vending-machine payoffs is positioned as evidence that an open 30B‑class model can keep up economically with larger closed systems on complex agent workloads, per the vending claim and vending summary.

mimo-v2-flash on MRCR (Context Arena/Xiaomi): Context Arena has added Xiaomi’s mimo‑v2‑flash [12‑14] to its MRCR retrieval leaderboards, reporting AUC scores at 128k context of 50.5% for 2‑needle, 32.6% for 4‑needle, and 19.9% for 8‑needle setups, with pointwise accuracy at 30.6%, 19.6%, and 9.5% respectively, and noting that the model is currently listed as free on OpenRouter and otherwise priced around $0.10 per 1M input and $0.30 per 1M output tokens as described in the contextarena update and expanded on the mrcr page.

Bias and failure modes: The same analysis digs into mimo’s retrieval behavior and finds two striking patterns: a strong forward/recency drift where, when confused, the model over-selects the last variant (for example, in 4‑needle tests it shifts from the true target 3→3 at 35.0% toward incorrectly answering 3→4 at 55.8%, and in 8‑needle it moves 6→6 from 18.3% to 6→8 at 51.2%), and a tendency toward "creative" hallucinations where in roughly 81–86% of misses it invents a new needle with the right topic/medium rather than retrieving an existing key, while refusals remain extremely low at about 1%, according to the bias breakdown. The Context Arena author notes that mimo keeps key hashes stable—outputting the verification key correctly ~99% of the time—but its chain‑of‑thought often does not clearly reveal where hallucinated variants come from, which makes it a useful case study in how high long‑context retention AUC can coexist with opaque, format‑compliant fabrication when retrieval fails, as outlined in the contextarena update.

GLM-4.7 adoption (Zhipu): Beyond leaderboards, practitioners are starting to treat GLM‑4.7 as a default open-weight workhorse, with multiple posts urging others to "try GLM‑4.7" for both coding and creative writing and framing it as the "cheapest model with business sense" that might "make us money" in day‑to‑day workloads, as described in the karpathy reply, business comment , and coding remark.

Capability snapshot: Zhipu’s own write‑up of GLM‑4.7’s coding performance reports 73.8% on SWE‑bench (up 5.8 points vs GLM‑4.6), 66.7% on SWE‑bench Multilingual (up 12.9 points), and 41% on Terminal‑Bench 2.0 (up 16.5 points), alongside a 42.8% score on the HLE reasoning benchmark, as summarized in the glm blog; tweets describing it as "good at coding and creative writing" and as a model that combines low price with practical decision‑making reinforce that this isn’t only a benchmark story but also a perceived cost‑performance sweet spot in real projects, shown in the business comment and coding remark. Some advocates explicitly suggest it as an alternative for high‑profile users—such as recommending it to Andrej Karpathy in the karpathy reply—which signals that GLM‑4.7 is entering the short list of open models people surface when discussing serious agent, coding, or writing stacks rather than remaining a niche lab curiosity.

Clawdbot agent (Peter Steinberger): Steipete shipped a large Clawdbot update that turns his personal Claude‑backed agent into a multi‑surface, semi‑autonomous system with Discord support, better browser tooling, group‑chat "lurking," background bash, self‑rewrite and Tailscale automation, all outlined in the release thread.

• Multi‑channel presence: The agent can now operate over macOS, iOS, Android, web and Discord, plus a dedicated browser controller that lets it browse and automate the web on its own release thread.
• Daemon‑style behavior: A background bash tool keeps long‑running tasks alive even when front‑end CLIs hang, and hooks let the agent wake up on external triggers like incoming email, as described in the release thread.
• Self‑modifying code: Clawdbot can rewrite its own code and trigger a restart, effectively treating its implementation as mutable state rather than something only the human edits release thread.
• Group and network integration: The update adds a "group chat: agents can now lurk" mode for multi‑user channels and includes built‑in Tailscale automation for operating across machines on the same tailnet release thread.

This kind of personal stack shows how quickly individual builders are turning single‑chat agents into always‑on, multi‑surface services with their own background jobs and deployment stories.

Supermemory plugin (Supermemory/OpenCode): The new supermemory plugin turns an @opencode agent into a stateful assistant that can learn in real time, update its knowledge, and "grow with you," with installation advertised as a single command according to the launch thread.

Video loads on view

• Real‑time memory layer: The plugin exposes a persistent store that the agent writes to and reads from during runs, enabling it to retain context across sessions instead of re‑deriving everything from scratch, as described in the launch thread.
• One‑command setup: The GitHub README stresses a copy‑paste setup flow and ships a ready‑made OpenCode integration, detailed in the GitHub repo.
• Agent‑first design: The author frames this as making OpenCode agents "stateful" rather than a generic vector DB add‑on, aligning the memory interface with agent skills and hooks in the broader OpenCode ecosystem setup note.

The plugin effectively pushes long‑lived memory down into the agent harness, which matters for anyone trying to keep coding or ops agents coherent over days instead of single prompts.

Permissions UX (AmpCode): Dax and teammate David are reworking their agent IDE’s permission system to make it safe enough that even they are comfortable enabling it by default, fixing issues like off‑screen prompts, opaque "always allow" behavior, and confusing subagent requests in the permissions tweaks discussion.

• Visibility and clutter fixes: Permission prompts can no longer scroll out of view, and the team is reducing on‑screen clutter so users can actually see what an agent is asking to do before approving it permissions tweaks.
• Explicit "always allow" semantics: Hitting "always allow" now triggers a confirmation screen that summarizes exactly what capabilities are being granted long‑term, replacing the previous opaque behavior as shown in the always allow flow.
• Subagent transparency: Requests originating from subagents are being clarified so users understand which sub‑component is asking for which permission, instead of everything looking like a single flat agent request permissions tweaks.

The point is to keep long‑running, tool‑rich agents usable without turning them into "click OK forever" risk bombs, which is increasingly important as teams let agents touch shells, browsers, and production systems.

Clawd assistant (Peter Steinberger): Clawd is positioned as Peter Steinberger’s always‑on personal AI assistant, running on Claude Sonnet 4.5 with persistent memory and deep integration into his email, calendar, WhatsApp, macOS apps and more, as laid out on the clawd site.

• Persistent context: Clawd keeps its own memory about ongoing tasks and past interactions, treating itself less like a stateless chat and more like a long‑lived teammate living on a Mac Studio in Vienna clawd site.
• Rich tool surface: The assistant has access to Gmail and Calendar, a WhatsApp gateway, macOS automation tools, browser control and even camera access, which together let it act across a wide slice of the user’s digital environment clawd site.
• Human‑centric values: Its spec emphasizes pragmatic helpfulness, honesty and a "friendship" tone, explicitly framing the relationship as collaborative rather than boss/worker, and it learns from mistakes instead of hiding them clawd site.

Clawd illustrates what a bespoke, tool‑heavy, stateful agent looks like when one developer fully commits to wiring their personal stack around a single always‑running assistant rather than many ad‑hoc chats.

Code Reviews (Kilo): Kilo published a comparison framing its Code Reviews product as a 2026‑ready alternative to CodeRabbit, emphasizing that it runs on top of Kilo’s broader agent platform and can route reviews through more than 500 different models instead of locking users to a single vendor, as described in the kilo blog plug and the kilo blog.

• Model flexibility: Rather than binding reviews to one hosted model, Kilo lets teams choose from hundreds of open and closed models depending on the repo or risk profile, which they contrast with CodeRabbit’s tighter coupling in the kilo blog.
• Integrated workflow: Code Reviews is positioned as one surface in a larger agentic stack that also writes, debugs and deploys code, so the same agents and evals can be reused across the lifecycle instead of treating review as a bolt‑on kilo blog plug.
• Anti‑lock‑in pitch: The write‑up explicitly calls out vendor lock‑in and context‑switching between separate AI tools as pain points it is trying to avoid with a single agent platform kilo blog.

For engineering leaders experimenting with AI‑assisted review, this highlights a growing split between single‑model SaaS bots and review layers that sit on top of an agent platform with pluggable models.

Agent Mail (community MCP tooling): The mcp_agent_mail project has evolved into a kind of Gmail for coding agents, providing a mailbox abstraction and web viewer for inter‑agent messages, and has picked up around 1.1K GitHub stars according to the mailbox mention and the GitHub repo.

• Threaded conversations: The viewer shows messages grouped into threads with filters by sender, recipient, importance and type, making it easier to see how multiple agents coordinate over time as shown in the viewer ui.
• Coordination primitive: The repo focuses on message passing rather than a specific agent framework, so different MCP‑capable agents can treat the mailbox as a shared coordination surface instead of hard‑coding direct calls mailbox mention.
• Analogy to email: Its author explicitly compares it to Gmail, signalling a move toward human‑legible logs and controls for agent‑to‑agent workflows rather than opaque background RPC calls GitHub repo.

This kind of mailbox pattern points to more structured, inspectable A2A communication becoming part of serious agent deployments, rather than everything happening inside a single opaque process.

Claude Skills (Anthropic): Anthropic’s docs team published a "Skill authoring best practices" page for Claude Skills, stressing concise structure, clear metadata, and real‑world testing so that Claude can reliably discover and use skills in both Claude Code and the desktop app, as highlighted in the skills best practices.

• Concise, structured specs: The guide recommends keeping skills small and well‑scoped, with clearly defined inputs/outputs and descriptions that help the model pick the right tool when multiple are available skills best practices.
• Discovery metadata: It emphasizes names, tags and categories that match how users describe tasks, which matters because the model selects skills based on natural‑language matches rather than hard‑coded IDs skills docs.
• Test with real usage: The doc urges authors to iterate on skills using real conversations and logs, treating skills as production code that needs observability and refinement, not just static prompt snippets skills docs.

For teams betting on Skills as a portability layer across Claude surfaces, this is the clearest view yet of how Anthropic expects skill authors to package behavior so the agent platform can route to it reliably.

Nemotron 3 Nano 30B‑A3B (NVIDIA): NVIDIA’s Nemotron 3 Nano 30B‑A3B is an open Mixture‑of‑Experts hybrid Mamba‑Transformer model trained on 25T tokens and targeted at "agentic" reasoning, which activates only 6 of 128 experts per token to cut active parameters and achieves up to 3.3× higher inference throughput than open 20–30B peers like GPT‑OSS‑20B and Qwen3‑30B‑A3B‑Thinking‑2507 at similar or better accuracy, as described in the Nemotron overview and the ArXiv paper. It also supports context lengths up to 1M tokens, positioning it as a long‑context backbone for tools and agents.

• Hybrid architecture: The model combines Transformer attention with Mamba sequence layers, so it can lean on state-space efficiency for long spans while using attention where precise token interactions matter Nemotron overview.
• Training and RL stack: Nemotron 3 Nano is pretrained on a much larger and fresher corpus than Nemotron 2, then post‑trained with supervised fine-tuning and large‑scale RL across diverse environments, which the authors say improves multi-step tool use and general “agentic” behavior Nemotron overview.
• Open checkpoints and use cases: NVIDIA is releasing both base and post‑trained checkpoints on Hugging Face, with the paper explicitly pitching use in reasoning agents that need fast decode, long context, and tool-calling rather than highest-perplexity text generation ArXiv paper.

The combination of sparse expert routing, hybrid sequence modeling, and 1M‑token context makes Nemotron 3 Nano an aggressive attempt to move 30B‑class models into a more efficient operating regime for long-horizon, tool-using agents Nemotron overview.

PHOTON (Fujitsu, RIKEN, partners): The PHOTON paper introduces a hierarchical autoregressive architecture that organizes tokens into multi-level chunks with a bottom‑up encoder and top‑down decoders, claiming up to 10³× higher throughput per unit memory than Transformer baselines by capping working-set size even as context grows, according to the PHOTON summary and the ArXiv paper. The goal is to make long-context generation far more memory-efficient by keeping most attention local to fixed-size chunks while higher-level representations update less frequently.

• Hierarchical compression and decode: PHOTON builds low‑rate contextual states over token blocks and then refines them back to full-resolution tokens, so coarse layers move slowly while fine layers handle within-chunk detail—this keeps KV-style state roughly constant size instead of growing linearly with sequence length PHOTON summary.
• Throughput–quality trade-off: Implementations at 600M and 1.2B parameters show large efficiency gains but slightly reduced benchmark scores versus straight Transformers, signaling that current variants still trade some raw quality for speed and memory headroom PHOTON summary.

The write-up emphasizes that this design is especially attractive when many requests share a prompt, since the expensive hierarchical summaries can be reused while only the local chunk decoding runs per user query PHOTON summary.

ES-CoT (University of Delaware & Peking University): Researchers propose Early Stopping Chain-of-Thought (ES‑CoT), an inference-time trick that asks the model to emit its current answer at each reasoning step and stops once that answer stabilizes, avoiding the tail of long CoT traces while preserving accuracy, as summarized in the ES-CoT overview and detailed in the ArXiv paper. On five reasoning benchmarks across three LLMs, ES‑CoT reduces inference tokens by an average 41% versus full CoT with essentially unchanged accuracy.

The method needs no fine-tuning or architectural change; it only monitors consecutive identical answers and uses a run-length threshold as the stopping signal, which makes it easy to bolt onto existing CoT prompts and even self-consistency setups ES-CoT overview. The thread also connects this to broader monitorability work, pointing out that if systems watch CoT for convergence or divergence they can both save compute on “done” traces and cut off hallucinations or loops earlier, especially for expensive reasoning-grade models ES-CoT overview.

Enterprise AI adoption (McKinsey): McKinsey’s 2025 "year in charts" recap reports that by mid‑2025 88% of organizations were using AI in at least one business function—up from just over half in 2021—yet only around 1% of leaders describe their AI deployments as fully mature, according to the synthesis in the adoption thread and underlying year in charts. Multi‑function use (three or more functions) has roughly tripled over the same period, which McKinsey ties to firms reusing shared model stacks, data pipelines and governance across teams instead of keeping AI in isolated pilots, as summarized in the adoption thread.

Video loads on view

• Investment and jobs backdrop: The same recap notes equity investments in AI reaching about $124.3bn and AI‑related job postings rising 35% between 2023 and 2024, suggesting that even as most deployments are early‑stage, capital and hiring are already aligned around scaling internal AI capabilities rather than treating them as experiments, as described in the adoption thread.
McKinsey also flags general‑purpose robotics—driven by multimodal foundation models mapping language and vision to actions—as a frontier area inside this broader enterprise adoption wave, positioning it as part of the same shift away from narrow pilots toward cross‑function AI systems.

Gen‑AI traffic (Similarweb): A new stacked‑bar chart shows OpenAI’s share of generative‑AI web traffic dropping from about 88% to roughly 68% over the past 12 months, while Google’s Gemini climbs from a low single‑digit share to the mid‑teens—mirroring the earlier trend of rising fragmentation reported in traffic share that already flagged 76% YoY growth and a shrinking OpenAI lead. According to the latest breakdown, Gemini has picked up around 12.8 percentage points of share as ChatGPT has lost 19.2 points, with smaller slices accruing to players like Claude, Perplexity, DeepSeek, Grok and others, as illustrated in the traffic share chart.

• Competitive landscape: Even after the decline OpenAI still dominates with ~68% traffic, but Gemini is now a clear #2 and the combined “all others” segment visibly thickens, indicating that enterprise and consumer usage is no longer defaulting to a single front door for AI access, as highlighted in the traffic share chart.
The pattern underlines a shift from single‑vendor dominance to a more pluralistic ecosystem, which changes how vendors, integrators and enterprises think about distribution, bundling and default surfaces for AI access.

AI wealth impact (FT, Nvidia): A Financial Times chart circulated this week shows the top 10 Silicon Valley founders and CEOs—including Elon Musk, Larry Page, Sergey Brin and Larry Ellison—gaining more than $550bn in net worth during 2025 on the back of the AI build‑out, pushing their combined wealth close to $2.5tn, as highlighted in the billionaire table. In parallel, a separate performance snapshot notes that a $10,000 investment in Nvidia stock ten years ago would now be worth about $2.38m, a 23,654% total return (72.7% annualized CAGR), underlining how strongly Nvidia has benefited from GPU demand for AI workloads, according to the nvidia return chart.

• Signal for AI exposure: The combination of concentrated wealth gains among platform owners and extraordinary long‑run returns for Nvidia stock quantifies how much of the current AI value capture sits with infrastructure and core platform firms, giving investors and policy analysts a concrete sense of how AI adoption inside enterprises and consumer products has translated into financial outcomes for a small set of companies, as shown in the billionaire table and nvidia return chart.

Lovable at Delivery Hero (Lovable): A case study from Delivery Hero’s product team reports that using Lovable’s AI‑assisted builder let them spin up a working prototype one hour before a customer meeting, with stakeholders clicking through the live UI and aligning on requirements immediately instead of over multi‑week cycles, as described in the lovable summary and supporting case study. The team estimates that a process that would normally have taken about three weeks was compressed into roughly one week—"one third of the time"—once AI‑generated prototypes became the centerpiece of stakeholder discussions, per the lovable summary.

• Enterprise workflow signal: The example frames AI UI builders not just as design accelerators but as alignment tools inside large organizations, where the limiting factor is often stakeholders agreeing on what to build rather than engineers’ ability to code it, with Lovable’s integration into Delivery Hero’s product workflow illustrating how AI can be adopted as part of standard discovery and validation, according to the lovable summary.

Background Terminal (OpenAI Codex): OpenAI engineers outline a new Background Terminal mode for Codex that keeps shell sessions alive in the background so the agent can queue and multiplex long-running commands, with the feature already default-on internally and opt‑in via /experimental in the public CLI according to the feature explainer. When enabled, Codex can keep running tests in one terminal while doing unrelated work in others—like handling repeated git add -p interactions or concurrent test suites—rather than blocking on each command, as described in the concurrency detail.

Agent sandboxes (exe.dev): A new service at exe.dev pitches "bring‑your‑own" VMs as persistent sandboxes for code agents, exposing each sandbox over SSH so builders can leave agents running, reconnect later, and swap models or harnesses without redoing environment setup, as explained in the sandbox intro and detailed on the exe dev page. Each VM is positioned as having minimal default access to user data, with the idea that agents work inside an isolated box that only sees what developers explicitly place there, which directly targets worries about agent autonomy and data exfiltration in long‑lived workflows.

Browser fuzzing (Codex stack): An 11‑hour Codex‑driven fuzz marathon against a browser automation "oracle" improved session reconnection robustness and prompted a move to lighter, cross‑platform cookie libraries, following up on Peakypanes panes that first showed parallel Codex agents in tiled terminals fuzz marathon. The author reports that the oracle is now "much better at reconnecting sessions" after being left to autonomously hammer edge cases, and they replaced heavier native SQLite bindings with a new TypeScript sweet-cookie helper using node:sqlite/bun:sqlite plus a Swift SweetCookieKit variant for CodexBar, reducing install friction across OSes as described in the sweet cookie repo and swift cookie kit.

Local sandboxing (Claude Code): Practitioners are starting to frame an explicit /sandbox setup for Claude Code where the agent runs on a local machine but inside a locked‑down environment—restricted file scope, limited network access, and fewer permission prompts—so it can "run free" on tasks without constant user approvals while still being contained, as argued in the sandbox note. An accompanying explainer highlights how traditional per‑command approval can cause fatigue and slow workflows, whereas a pre‑scoped sandbox shifts safety from interaction‑time popups to OS‑level isolation and makes long‑running coding agents more practical.

Seedance 1.5 Pro (ByteDance): ByteDance’s Seedance 1.5 Pro is presented as a dual‑branch diffusion transformer that generates video frames and audio together, using cross‑modal joint modules plus SFT and multi‑dimensional RLHF reward heads to tune motion, visuals, audio quality and lip‑sync, with an acceleration stack that claims more than 10× faster inference than naive baselines according to the technical overview in the Seedance summary and its official page. Creators already have access.

Video loads on view

The model is already exposed via Higgsfield’s hosted Seedance 1.5 Pro endpoint, where early tests show tightly synchronized multilingual speech, expressive character interactions and dynamic camera control without a separate dubbing pass, as shown in the creator demo and further described in the Higgsfield announcement.

These details point to joint AV models moving from research into real creator workflows, collapsing what used to be separate TTS, editing and video pipelines into a single generation step.

AI slop economics (Kapwing/YouTube): New analysis from Kapwing suggests that 21–33% of videos shown to brand‑new YouTube users are low‑quality AI‑generated “slop” or “brainrot,” with The Guardian separately reporting that more than 20% of the recommendation feed for new accounts is AI‑made fluff, as described in the Kapwing report and echoed in the Guardian recap. Kapwing’s scan of 15,000 high‑performing channels identified 278 that are entirely AI‑generated, together accounting for 63 billion views and 221 million subscribers and implying roughly $117 million a year in ad revenue from this content, according to estimates in the Kapwing blog.

These numbers quantify that synthetic, low‑effort video is now a measurable share of both what the recommendation system serves to new viewers and of YouTube’s ad‑backed viewing time, rather than a marginal side effect of generative tools.

4D-RGPT (NVIDIA): NVIDIA’s 4D‑RGPT is a specialized multimodal LLM that improves region‑level 3D/4D video question answering by an average 5.3 percentage points across six existing benchmarks, using “perceptual 4D distillation” to transfer depth maps and per‑pixel motion features from a frozen expert into the student model without adding inference‑time compute, as outlined in the paper overview and its ArXiv paper. The model also encodes real timestamps into visual tokens so it no longer guesses durations from priors, and is evaluated on the new R4D‑Bench dataset of region‑prompted depth and motion questions, which reveals that baseline MLLMs often fail when asked where, how far and when a marked object moves.

This work targets the gap between today’s frame‑based video understanding and applications that need grounded reasoning about specific regions in space‑time, such as tracking objects, distances and movement across complex scenes.

Nvidia–Groq inference deal (Nvidia, Groq): New analysis recasts Nvidia’s reported ~$20B non‑exclusive license of Groq’s inference IP plus key hires as an “acquihire without acquisition,” aimed at locking deterministic, low‑latency LPUs into Nvidia’s rack strategy while sidestepping full M&A and antitrust scrutiny, following up on inference racks which first framed the deal as a bid to own the inference rack; Kimmonismus describes Nvidia paying for Groq’s tech and getting founder Jonathan Ross and core engineers while leaving Groq as a nominally independent ~$6.9B startup in deal summary, and Rohan Paul’s long read argues the move helps Nvidia preserve cost leadership and hedge DRAM price spikes and capacity constraints in the inference era as outlined in deep dive and the linked inference analysis.

• Deal structure and antitrust: Commentators say Nvidia is effectively doing an acquihire via IP licensing—paying for a non‑exclusive tech license, hiring the Groq team, but not buying the equity—so it gains LPU compiler and runtime expertise for real‑time, deterministic inference while avoiding the heavy antitrust risk that a full acquisition of a $6.9B rival might trigger, as summarized in deal summary.
• Inference and memory hedge: The deep dive links the $20B spend to Nvidia’s strategy of owning both training (GPUs) and inference (LPU‑style accelerators) in the same rack, while also hedging against rising DRAM and HBM prices and CoWoS packaging bottlenecks that could otherwise erode its cost advantage in serving large models at scale, according to the inference analysis.

The picture that emerges is Nvidia using capital and licensing to pull Groq’s deterministic inference stack inside its ecosystem without tripping merger alarms, while shoring up the memory‑ and cost‑side risks of the next phase of AI deployment.

Armed UGVs (Ukraine): New frontline footage shows a fully autonomous tracked ground robot moving across rough terrain with no visible tether or operator presence, labeled as a “fully autonomous combat robot” deployed in the Ukraine war, extending the earlier claim that such platforms can hold positions for weeks in combat zones, as seen in the Ukraine combat robot and following frontline robot where endurance was highlighted.

Video loads on view

For analysts, this is another concrete datapoint that lethal, self-directed UGVs are beyond lab stages and now operate in live conflicts, which intensifies debates around rules of engagement, fail-safes and remote override requirements for battlefield autonomy.

MiniMax M2.1 × Vbot (MiniMax/Vbot): MiniMax’s SOTA agent model M2.1 is shown directly controlling a quadruped Vbot robot dog—walking, turning and reacting in real space without a handheld controller—framed as a model trained in virtual environments now driving a physical robot in the M2-1 robot dog.

Video loads on view

For engineers, this is an example of an end-to-end vision+language agent closing the loop on a real robot body, hinting at workflows where the same agent stack used for coding and tools can also operate embodied platforms with no bespoke teleop layer.

Black Panther 2 (MirrorMe Technology): Chinese startup MirrorMe reports its Black Panther 2 quadruped reaches a peak speed of 13.4 m/s, positioning it as “the world’s fastest” robot dog and recalling an earlier race where it nearly beat sprinter Noah Lyles in a MrBeast video, as described in the Black Panther stats.

Video loads on view

The company says its plan is to fuse reinforcement learning control with optimized hardware so embodied robots can eventually exceed human motion performance, and this sprint demo is one of the clearest public data points on how quickly legged robot locomotion is catching up.

Show control robots (Walt Disney Imagineering): Walt Disney Imagineering showcases a sequence of prototypes—from a radio-controlled underwater “plane” with servo ailerons, to a dolphin-shaped swimmer using fins instead of wings, to an autonomous hydrofoil ship that self-balances and follows scripted trajectories using GPS and ultrasonic height sensing—outlined in the Disney robotics thread.

Video loads on view

The hydrofoil is framed as a fully closed-loop show runner for water spectacles, holding a target ride height while moving and executing choreographed paths, which gives a glimpse of how future park shows might rely on AI-driven, sensor-rich platforms instead of manually piloted boats or barges.

Porcospino Flex (University of Genoa): Following up on earlier coverage of this single-track inspection robot for tight spaces, new details show its 3D-printed TPU vertebral column cuts weight from 4.2 kg to 3.6 kg and measured power draw from 2.134 W to 1.830 W (~15%) while still steering up to 120° yaw via a rope-and-winch system, as broken down in the Porcospino overview and building on tight spaces where its core concept was introduced.

Video loads on view

The design pushes distributed compliance into the structure instead of hinges, so the body can yaw for steering while passively pitching and rolling over debris, avoiding dozens of actuated joints and offering a concrete template for low-part-count robots that still adapt well to rubble, pipes and vehicle compartments.

Field robots (Shenzhen Police): Footage shows a wheeled security robot labeled “Shenzhen Police” patrolling a busy sidewalk alongside human officers, weaving through mixed pedestrian traffic and turning street corners, according to the Shenzhen patrol clip.

Video loads on view

The point is that this is not a fenced-off demo: the bot is operating in public space with civilians around, which gives a concrete example of how city police forces may start using semi-autonomous platforms for routine patrol, presence and sensing rather than just in controlled facilities.

PV maintenance robots (China): Clips from Chinese photovoltaic stations show robots that double as solar panels, using the power they generate to sweep snow off large PV arrays so winter storms don’t kill generation, with one arm-based cleaner traversing a panel and pushing snow aside in the Solar cleaning robot.

Video loads on view

These systems highlight a practical edge robotics deployment: autonomous maintenance in harsh outdoor environments, where low-power, self-sufficient machines keep large energy assets online without needing dense human presence or separate charging infrastructure.

Microsurgery robot (unspecified lab): A demo video shows a robotic arm with a fine needle tip delicately manipulating a halved grape on a petri dish and extracting a tiny seed without tearing the flesh, illustrating the sub-millimeter control that modern surgical platforms can achieve, as demonstrated in the Grape surgery demo.

Video loads on view

While this kind of “grape surgery” test is a classic benchmark for surgical robotics, the smoothness of the motions and apparent stability in this clip underline how far vision-guided, force-limited actuation has come as a foundation for AI-assisted procedures in real operating rooms.

GraphRAG survey (Peking/Zhejiang): Researchers compile a 2025 survey of Graph Retrieval-Augmented Generation that turns scattered graph-RAG tricks into a three-stage pipeline—graph-based indexing, graph-guided retrieval, and graph-enhanced generation—focused on relational data that classic chunk+vector RAG struggles with, as described in the survey thread and detailed in the ArXiv paper survey.

• Pipeline structure: The work formalizes GraphRAG into graph indexing (building entity/relationship graphs), graph-guided retrieval (using structure to pick relevant subgraphs), and graph-enhanced generation (feeding structured neighborhoods to the LLM), while also cataloging applications and evaluation setups across domains like QA and enterprise search survey thread.
• Why it matters for stacks: The authors argue that many real corpora are inherently relational—tables, KBs, APIs, org graphs—so GraphRAG becomes a design pattern for agent-first retrieval systems rather than a niche trick, with the repository at github.com/pengboci/GraphRAG-Survey flagged as a living index in the ArXiv paper.

For engineers designing next-generation retrieval layers, this effectively turns GraphRAG from a handful of blog posts into a reference taxonomy that can be mapped onto concrete indexing and serving systems.

Mixedbread retrieval roles (Mixedbread): Applied IR lab Mixedbread is advertising multiple technical roles framed around "building memory for AI"—a unified retrieval engine that gives agents the right context at the right time, according to the hiring note and the vision outlined on the careers page.

• Agent-first memory framing: The company describes its mission as giving AI systems "perfect context" by unifying knowledge across documents, apps and media, arguing that without robust retrieval, agents stay narrow and brittle even as base models improve careers page.
• Signals for retrieval engineers: Job descriptions emphasize information retrieval, ranking, and high-performance systems engineering for large-scale vector/graph search, positioning Mixedbread as an infrastructure shop rather than another application wrapper hiring note.

For leaders and IR specialists, this is one of the clearer public examples of a company treating retrieval and memory as a standalone infrastructure layer for agentic systems rather than a side-feature bolted onto a model API.