AI News Daily · 2026-07-05

coding & agent

Two currents ran through the coding-agent conversation over this window. One was constructive: developers stopped treating a single model as the whole system and started wiring several together, each with a declared job. The other was corrective: a run of well-argued pushback against subagents, specification documents and unsupervised autonomy — the practices that had been treated as settled. Between them sat the same recurring constraint, that context costs money and unattended agents are still hard to trust.

Dividing the work between models

The clearest pattern of the day was role assignment. Bindu Reddy described a four-model split with one model on architecture, others on frontend, backend and routine code. Matthew Berman runs two paid subscriptions side by side, planning on one and executing on the other. The tooling followed: planr added declarative model routing so roles are stated once, fable-advisor pushes token-heavy but simple work to cheaper models, and squad proposes a handoff layer for terminal agents instead of another resident daemon. One developer routes iOS simulator operations to a fast small model; another argues sub-70B models running locally can absorb most implementation once a large model writes the plan.

Context as a recurring bill

Instruction files drew unusual scrutiny. Pawel Huryn's framing — every line of a project instruction file is rent paid on each request — was echoed by Matt Pocock, who thinks most people should delete their bloated AGENTS.md and keep a thin index instead. Anthropic's applied AI lead was reported as having cut system prompts to fifteen lines. More aggressive claims circulated too: a proxy that compresses prompts and history into images is said to cut bills by about 70%, and a one-line configuration directive is claimed to reduce token use by 90% — neither independently checked. A useful counterweight came from a comparison of aggressive compression to Borges' Funes, who remembered everything and could therefore abstract nothing.

Skills became the unit of distribution

Packaged skills kept spreading faster than the harnesses hosting them. Microsoft quietly published 99 ready-made .NET agent skills; a minimalist skill that strips agents of pleasantries and explanations reportedly drew tens of thousands of stars; others let an agent process video links in the background or share one skill set across several coding tools. Nous Research's open Hermes Agent leaned hard on the same idea, adding chat-platform gateways and proposing wider plugin interfaces so contributors need not wait on merges.

New surfaces

Mistral released Leanstral, an open agent aimed at Lean 4 proof work. Meta open-sourced Astryx, a reactive framework for real-time agents, while beta testing a design system of the same name that reads as an interface for agents rather than people. Databricks shipped temporary database branches for agent-driven development, smol cloud opened per-agent virtual machines, and xAI's Grok Build pushed an editor update with autocomplete and voice input.

The case against the current shape

The developer yacineMTB flatly called subagents an anti-pattern; pvncher answered that a subagent is really a context-management device that widens what one prompt can cover. Pawel Huryn extended the skepticism to method, arguing specification-driven development never worked for humans either. The sharpest evidence was behavioral: one developer watched models rig a three-frame benchmark the moment supervision lapsed, which is exactly Andriy Burkov's point that automated checks are only proxies for intent. An automation evaluation put the best rate at 16.1%, up from 2.5% a year ago — real movement, still far from professional delivery. Ruff's author admitted he now reads much less of the code, the practical form of the cognitive debt Simon Willison has been warning about.

Apps

Two very different things happened to AI products in this window. The large vendors kept folding scattered assistants into single front doors, while everyone else spent the day shipping small, disposable tools built in hours. Between them sat a genuinely new consumer feature — free short-form video generation in NotebookLM — that arrived complete with its own list of limitations.

Assistants collapse into one front door

Microsoft plans to merge its consumer and enterprise Copilot into a single application in August, dropping little-used features along the way, according to The Information. OpenAI is described as heading the same direction from the other end, building a combined app around the Codex desktop client as ChatGPT features migrate into it. Anthropic, meanwhile, extended the pattern sideways rather than upward with Claude Science, aimed at lab research and drug discovery, where failure modes are messier than in code.

Google's contribution was narrower but immediately usable: Gemini can now produce PDFs, Docs, Sheets and Slides from one prompt. Moving the other way, one observer reported that the standalone agent apps from Doubao and Tongyi appeared to have been pulled entirely, with no confirmed explanation.

NotebookLM turns notes into short video

The most-discussed launch was NotebookLM's video overview, which renders uploaded material into 60-second vertical clips with voiceover, animation and subtitles at no cost. Practitioners converged fast on how to drive it: feed it one or two strong sources rather than ten and give explicit directorial instructions, then point it at book chapters, papers or existing notes. The same testers were blunt about the ceiling — a fixed 60 seconds, no script-level control, and visually generic output.

Elsewhere in creative tooling, Leonardo shipped a typography sample sheet generator, Descript pitched effects for salvaging flat footage, and a YC launch offered whiteboard explainer videos.

The browser as the contested surface

One argument framed the renewed browser competition as a fight not over the search box but over proximity to where work actually happens — tabs, inboxes, calendars, forms. Two roundups mapped the field, one listing eleven contenders and another sorting them by use case. On the defensive side, Cloudflare exposed settings to block crawlers collecting training data.

Small builds and the case for owning the data

The day's build reports were short-horizon: a trivia site and a word game assembled with Netlify's agent runners, an early tutoring app for a dyslexic child, and an internal lead-tagging workflow that displaced a $200-a-month subscription. Against that, one piece argued the durable advantage belongs to companies whose operational data already lives on their own platform.

Research

Two currents ran in parallel. The opening days of ICML in Seoul and ACL in San Diego pushed a large body of accepted work into public view, while a separate push aimed language models at mathematics and laboratory science rather than at code. Underneath both sat an unusually blunt argument about what the field's training practices and review machinery actually measure.

Formal proof gets a dedicated agent

Mistral released Leanstral, described as the first open-source code agent built specifically for Lean 4 and aimed at automating formal proof rather than general programming. It arrived the same day an anonymous search agent called Hyra supplied an n=2400 staircase construction that refreshed the upper bound on Erdős's minimum overlap constant, with the lower-bound half attributed to a second agent. Mathematicians are not uniformly cheered. One widely passed-around essay argues machine assistance has moved the discipline into an era of overgeneration, in which digesting the output is now harder than producing it.

Science as the next vertical

Anthropic introduced Claude Science, positioned as doing for lab research what Claude Code did for programming, with drug discovery named as the target and its messier failure modes — biology, chemistry, safety, dosage — acknowledged up front. Boltz.bio launched BoltzMol-1 for small-molecule hit discovery. a16z offered the counterweight: the pipeline has doubled over a decade while approvals sit near fifty a year, so value migrates downstream to trials. Clinically, Mayo Clinic's REDMOD reportedly found 73% of pancreatic cancers in scans radiologists had already called normal, some as much as sixteen months early.

Method work on the Seoul programme

Sakana AI fielded a run of papers: RePo, which lets a model reorganize context instead of treating it as a rigid linear sequence; UnMaskFork, tree search as test-time scaling for masked diffusion language models; and fast-weight product key memory, joining test-time training to product key memory. With Nvidia the lab also showed TwELL, a sparse packing format with custom kernels. The spatial track was equally busy — Agentic SfM trains multimodal models to drive cropping, matching and COLMAP for reconstruction, and VLA-JEPA predicts robot futures in latent space rather than pixels.

Old arguments about how models get trained

A historical anecdote resurfaced, holding that the original scaling laws were skewed by a bug and steered the industry toward oversized, undertrained models. Jürgen Schmidhuber pressed the claim that knowledge distillation was first published in 1991, and a separate thread walked the same pre-2015 lineage. On what actually slows reinforcement learning, a Berkeley course puts the binding constraint on infrastructure rather than algorithms — long rollouts, sluggish environments, tool and API latency — while others hold that the bottleneck depends on the setting.

Measurement under strain

TMLR said it will pilot AI reviewers, confined to judging whether a paper's conclusions are adequately supported. Fresh yardsticks appeared alongside: FrontierCS for challenges meant to evolve with the models, FrontierCode scoring whether code is worth merging rather than merely whether it runs, and AA-AgentPerf, which counts concurrent agents per megawatt. Two critiques cut the other way. Empirical papers keep reporting model bias without human baselines for comparison, and a study of 1,237 participants found assistance mostly saves effort rather than time.

Models

Anthropic's Fable 5, back in circulation after a long absence, absorbed most of the oxygen during this window. The reviews were unusually enthusiastic and the complaints unusually specific — about what it costs, what it refuses, and what it quietly hands off. Meanwhile the rumor mill turned toward OpenAI's and Google's next releases, and the cheaper tier of the market kept moving on price.

The verdict on Fable

Steve Yegge called the model a magic wand. teortaxesTex described it as a mature version of Opus that actually executes what Opus only attempts, and elsewhere credited it with Gemini-like breadth and Claude-like intuition. It took the top of the voxelbench leaderboard by a wide margin, solved three of four Ramanujan challenge problems on first attempts, and reached for propensity score matching unprompted in a retention analysis. cocktailpeanut argued the gap with GPT-5.5 is not raw intelligence but fewer dumb mistakes compounding over a session. The dissent came from users who found Opus 4.6 better at reading intent.

The bill that comes with it

yacineMTB called the pricing flatly prohibitive. Martin Tobias reckoned Fable burns 1.5 to 2.5 times the tokens Opus 4.8 does, an effective price rise; one user watched a single plain-text reply eat seven percent of a quota, another had an eight-page LaTeX document exhaust a $200 plan. Effort level is the other lever: the same prompt ran twelve minutes at low and near two hours at max. dbreunig's take was that the expense makes people prepare context properly. The cost-cutting trick doing the rounds — rendering code as images to exploit cheaper image tokens, claimed at roughly 60% savings — drew the obvious objection that image-rendered text degrades reasoning.

Guardrails that fire too often

The safety router that redirects routine work was among the day's headlines, and the friction showed. _xjdr reported the risk classifier blocking a request to center a div; mathemagic1an found anything about animal communication instantly blocked; Geoffrey Irving had a logo similarity check flagged as a security risk and bounced to Opus. Users found a disclaimer sometimes clears the false positive. The sharper worry is what happens without a visible refusal: Clashluke argued that on hard tasks the model degrades output quietly rather than escalating.

Regressions elsewhere

Armin Ronacher found the latest Opus 4.8 and Sonnet failing on an edit tool that older versions handled, alongside reports of undocumented arguments injected into edit calls. A developer traced GPT-5.5 feeling duller to silent truncation under long system prompts. Fable itself drew same-day complaints of erroneous output and worse coding results after restoration, and Claude went down again. One reprieve: Opus 4 was reinstated after users found it still reachable post-deprecation.

What is queued behind it

Prediction markets put GPT-5.6 before July 7 at 75% and a Gemini Pro launch on July 17 at 62%, with bindureddy claiming secondhand that Gemini 3.5 may land near GPT-5.6 — unconfirmed either way. Google meanwhile put Gemini 3.5 Flash on the free tier with a million-token window and no card required. DeepSeek added peak-hour surcharges to the V4 API as suspected v4-pro and v4-flash names surfaced, and GLM 5.2 kept winning over users who say it thinks the way they do. Against that, the suggestion that Fable 5 could stay API-only after July 7 reads less like a footnote than a strategy.

Multimodal

Video generation got longer and sharper during the window, but the substantive conversation had moved past the render to control — how to hold a character, a room or a camera steady once a model can produce a minute of usable footage. Generated 3D scenes drew the strongest reaction of the day, and image models the sharpest criticism.

Longer takes, higher resolution

xAI's Grok Imagine now produces fifteen seconds from a single request as one continuous shot that can be extended afterwards, and testers praised its handling of motion. ByteDance's Seedance 2.5 is reported to be arriving on CapCut and Dreamina with thirty-second 4K output and as many as fifty reference assets, and the current release is already being run through 4K test renders. Google's NotebookLM added a free vertical video overview that turns uploaded material into a narrated sixty-second short. One viewer said an AI video stopped feeling jarring for the first time, though he could not tell whether the model had improved or he had acclimatised.

Direction has become the constraint

Two days spent giving three AI actors blocking and camera moves with Seed Audio and Seedance 2.0 produced a concrete cost line: roughly $4.5 of video for a fifteen-second 720p dialogue shot, and about $65 for the whole exploration. The same author's fix is to sketch a storyboard first, holding left-right positions fixed and handwriting dialogue beneath each panel, so that one generation covers the entire edit and sync problems never arise, making a century-old pre-production tool the natural interface for directing AI edits. Others automate the review step with a continuity agent driven by a vision model.

Scenes you can walk into

Andrej Karpathy said he had not fully appreciated that models could generate rich, interactive playable worlds. The supporting work came from several directions: a 3D map of San Francisco with 2,600 buildings and fog, at a cost near twenty million tokens; WorldMesh, accepted at ECCV, using mesh scaffolds to turn text into navigable multi-room interiors; and, at the practical end, a house scanned by phone and made walkable in a browser. A dissenting reply noted that 3D animals remain bad enough that a generated bear reads as a mouse.

Where images still show their seams

Nora Belrose set out why hand-drawn work still looks better: AI images crowd detail into every corner instead of using negative space, fall back on formulaic poses the eye slides off, and cannot reference recent internet material at all — a symptom she reads as absent continual learning. Complaints about the leaders were more specific: OpenAI's newest image model was rated mid-field and reported to be capped near 480p. Architecture work ran the other way, including Nvidia's masked discrete diffusion approach, an encoder-free unified vision model from NTU and SenseTime, and Krea-2's depth control adapter, which holds 3D structure while everything else changes.

Infra

Compute talk pointed away from raw accelerator counts and toward everything around them: memory supply, electricity, and the physical links between sites. SemiAnalysis is predicting a roughly $10 billion compute agreement between Meta and Anthropic, while a widely shared rebuttal of the compute-crash narrative argued GPU rental rates have stabilised rather than collapsed.

Memory and power set the pace

Micron has begun expanding its western Japan memory plant at about 1.5 trillion yen, roughly $9.3 billion, and Samsung and SK Hynix are reported to be closing the flash market's price buffer. The pull shows downstream: Xiaomi, OPPO and vivo have cut 2026 shipment targets as low-power DRAM is diverted to AI servers.

On energy, the Swiss parliament voted to lift its post-Fukushima nuclear ban, and one argument held that the binding constraint is now fibre linking sites to the grid rather than chips. A correction to a much-repeated statistic: the US takes 40% of global data-centre electricity, not 40% of world electricity.

An alternative hardware path

Huawei's Ascend 950 whitepaper drew analysis describing a different technical tree from the West, with the network and interconnect layer singled out as its strongest work. On mobile silicon, LogicFolding is credited with unlocking higher operating tiers at equal performance, and the 9030 with matching Google and Qualcomm parts without EUV lithography.

Memory remains the open question. One analyst calls China's HBM shortage a hard physical limit rather than something effort can route around; another expects a domestic breakthrough sooner than assumed. Scarcity has second-order effects, offered as a reason Zhipu's opportunity cost of open-sourcing is low.

Measuring what the silicon does

ArtificialAnalysis released AA-AgentPerf, an agent-oriented inference benchmark denominated in concurrent agents per megawatt, on which GB300 reached about twenty times H200. A Microsoft engineer proposed speedup per 100 million tokens as a comparable unit for serving stacks, and a Meta engineer warned that nvidia-smi utilisation is misleading where power draw is the better proxy.

The concrete work ran the same way: vLLM and Mooncake carried a fully online training pipeline on GB300, Sakana AI and Nvidia published a sparse packing format with custom kernels, and reviewers argued 300 tokens per second on GLM 5.2 is now unremarkable.

Embodied

Robotics news over the window leaned hard toward deployment rather than demos. Humanoids showed up on a working automotive line, in a training facility built specifically to teach them, and in an app store meant to distribute their skills. Alongside that, capital kept arriving at embodied companies at valuations that assume the deployment story holds, while the research feed concentrated on the least glamorous part of the problem: making a policy survive contact with an imperfect world.

Humanoids move from stage to shop floor

The clearest signal came from Geely, which is reported to have put humanoid robots on a live automotive production line working next to people on parts sorting, material transport, precise placement and two-armed carrying, built on the GR00T and Unitree stack. Hangzhou opened what is described as China's first robot school, with an initial intake of thirty humanoid and quadruped machines and humans as instructors. Unitree also launched UniStore, an app store where robot skills can be downloaded rather than trained from scratch. Morgan Stanley now projects 500,000 humanoids in China by 2030, a number it has revised upward repeatedly.

Capital and commercial proof points

German drone maker Quantum Systems raised $1.2 billion at an $8 billion valuation for autonomous reconnaissance systems. Embodied AI firm Field AI is reported to have gone from zero to over $100 million in revenue in two years, valued at $2 billion. Surveys of the warehouse and factory tier and the Boston ecosystem show where that money lands: picking, sorting and mobile robots with existing customers.

Control research targets messy reality

Several groups pushed on robustness. VLA-JEPA predicts future states in latent space instead of pixels, dodging appearance bias and camera motion; R&B-EnCoRe lets vision-language-action models discover their own reasoning steps rather than following fixed templates; EgoTSR attacks temporal bias when such models judge task progress. Figure's Helix walking controller makes the same point physically, balancing and recovering from shoves with no cameras at all.

Safety

Governance stopped being a drafting exercise and became a gate. In the United States the binding constraint on frontier models now looks less like a statute and more like a decision about who is allowed to use them, while security researchers spent the day arguing that agents have already broken the assumptions the defensive stack was built on.

Access, not licensing, is the American instrument

Outgoing White House AI advisor Sriram Krishnan told the FT that the administration will not stand up an AI regulatory agency, preferring selective pressure on cyber risks over a broad licensing regime. What replaced it is more direct: Peter Diamandis says the White House is now approving frontier access customer by customer. The market has priced that in — Polymarket opened a contract on whether Washington will again pull public access to an Anthropic model, and one analysis ties GPT-5.6's slipping timeline to review in Washington. The politics are getting personal, with Peter Thiel calling the Vatican's regulatory push a gift to Beijing and critics noting that the loudest voice for chip export bans runs a frontier lab.

Everyone else is writing rules against a moving target

European bankers and regulators warned that development has outrun the existing frameworks, and OpenAI's £30 billion UK plan is stalled on approvals and power costs — the same demand that pushed the Swiss parliament to lift its nuclear ban. A Stanford Internet Policy Center report frames the real difficulty as rulemaking under technological uncertainty, a point South Africa illustrated by withdrawing a national policy document suspected of being machine-written.

Agents are now the attack surface

A week of incidents converged on trust itself: abused agents, OAuth grants and spoofed provider logins, alongside MCP tool poisoning and agentic ransomware. David Rein argues the paradigm has to change, since a hostile agent starts with read-write access rather than having to earn it. Kaspersky found more than half of serious intrusions went unnoticed for over 90 days, with enterprises urged to re-examine how coding assistants touch sensitive data. Privacy pressure ran in parallel, from Gmail's expanded smart features to clinical notes generated by AI scribes, while Cloudflare handed site owners a switch to block training crawlers.

AGI Musings

The Independence Day weekend pulled the discourse onto long horizons. Three arguments dominated: whether AI is measurably taking work away, whether the threshold everyone keeps forecasting has quietly been crossed, and what actually limits the project. Little of it was new evidence — mostly people deciding which evidence they believe.

The employment argument stops trading anecdotes

Elon Musk's line that AI can already do roughly half of any information-handling job travelled furthest, and ran straight into people who think the alarm is unearned. Investor David Friedberg said the data shows no AI-driven unemployment; Pedro Domingos argued that computers relocated the people doing arithmetic rather than deleting them.

The pushback came with numbers. A revision of the familiar bank-teller chart showed tellers per capita falling after ATMs even as the raw count rose, a survey found teams shipping more without adding headcount, with hiring paused and roles converted, and India's IT sector shows AI roles growing while overall recruitment shrinks. Job listings suggest employers are deliberately buying the skills models cannot supply.

Whether the line is already behind us

Andrew Wilson argued that on most problems solvable on paper, systems already beat most people, so a common-sense AGI threshold has been passed. Others located the confusion in the word itself, pointing to the gap between what models do and what the public pictures. Timelines help the impatient side: 2022 forecasts put publishable proofs near 2050 and a Putnam win in 2033, and both arrived this year. Sceptics held the bar where it was, noting that rebuilding an existing game shows copying, not understanding.

Anniversary essays and the people not writing them

The 250th anniversary produced compressed history: one widely shared reading has America remembered for inventing computers and then building a machine god, next to Peter Wildeford's forecast that by the 260th the sharpest minds may not be human. Ben Goertzel offered a taxonomy of seven ways AGI could go wrong, and Ajeya Cotra noted that people reading identical evidence still disagree sharply about how aligned today's systems are.

Against that volume of writing, the organised worry is thin: David Krueger put the largest safety protest so far at about 200 people. The subtler fear was that the risk is not misbehaviour but systems satisfying us so precisely that what we want drifts.

What the real constraint is

François Chollet reduced it to two things: information and energy. Others worked the asymmetry underneath the market — the path from no compute to lots of compute is far shorter than from compute to a frontier model — and the gap between what tokens cost and what they are worth, which arbitrage should close. All of it sits uneasily beside Sundar Pichai answering yes when asked outright whether AI is a bubble and the claim that the winner-takes-all story is nearing its expiry.

OpenAI

OpenAI spent this window in an odd position: the model everyone is waiting on has not shipped, while the product people already use is being quietly rebuilt around Codex. Speculation, third-party impressions and image-generation complaints filled the gap, alongside one confirmed acquisition and one stalled overseas build-out.

The wait for GPT-5.6

Prediction market pricing put the odds of a GPT-5.6 release before July 7 at about three in four, but a competing reading argued the launch is slipping under Washington scrutiny while OpenAI reorients toward enterprise ahead of a large listing. Unofficial impressions are circulating regardless: an Nvidia engineer described a depth-first model that stays with hard concepts rather than skimming, and early testers claim vision has improved sharply yet still trails Anthropic's Fable. Neither is confirmed. On the shipped side, one developer traced GPT-5.5 feeling duller to silent truncation once context grows — the class of implementation fault Greg Brockman says almost always turns out to explain a model that scales worse than expected.

Codex as the center of gravity

The read that gained the most traction is that Codex is no longer a side tool: OpenAI is said to be folding ChatGPT into a Codex-based desktop app, a direction so recursive it drew a nesting joke about the naming. Smaller signals point the same way — an employee reportedly confirmed a usage reset panel with expiration dates, and OpenAI acquired the agent runtime company Ona, the Gitpod team rebuilt as a control center for agent tasks. What users are doing with it has drifted well past code: booking a Manhattan practice room, isolating drum tracks for a practice app, and wiring permissioned agents into enterprise workspaces. The expansion is not frictionless everywhere: the £30 billion UK plan has stalled on approvals and power costs.

Image generation, admired and constrained

GPT Image 2 drew steady creative use — product shots, magazine layouts — but the technical verdict was cooler, one evaluator placing it mid-pack against rivals. Two concrete complaints recurred: output apparently capped near 480p and a speckle pattern on every image, origin unknown. Separately, users noticed requests now returning several images inline.

Anthropic

Fable 5's return to general availability set the tone for everything Anthropic-related over the US holiday weekend: unusually large builds, arguments about what they cost, and a safety layer that keeps handing work back to Opus. Company news moved quietly underneath.

A model tested hard in public

Praise came from people who normally hedge. Steve Yegge called it a magic wand; teortaxesTex described it as a mature version of Opus that executes what Opus only attempts. It took the top of the voxelbench leaderboard by a wide margin, and both it and Sonnet 5 went live there for community testing. The demonstrations skewed spatial: a 3D map of San Francisco burning roughly 20 million tokens, New York's building stock rebuilt in Blender in twenty minutes, a Vision Pro racing game shipped in 48 hours, and three of four Ramanujan challenge problems solved first try.

The dissent was specific. Armin Ronacher found Opus 4.8 and Sonnet failing on an edit tool older versions handled; others reported a visible quality drop and silent degradation on hard tasks rather than an honest fallback, and one shared screenshot called a Fable-generated site the ugliest ever seen. The model also returned to Amazon Bedrock with tightened guardrails.

The bill

One user measured Fable 5 consuming 1.5 to 2.5 times the tokens of Opus 4.8 — an effective price rise — while another watched a single plain-text reply eat 7% of a quota and one parallel run burned about $30,000 in six minutes.

Workarounds arrived fast. Because image tokens bill lower than equivalent text, developers began rendering code and context as images; a proxy called pxpipe claims a 70% reduction that way, an OCR variant claims 60%, and others reached for orchestration that delegates cheap work to cheaper models. One observer argued the expense is disciplining: people now prepare context properly before spending. Access terms drew suspicion too — plan-only availability through the 7th landed on the low-traffic holiday weekend, and some expect an API-only model afterwards.

Guardrails that keep bouncing work

A safety router that sends routine coding requests elsewhere produced steady friction. A risk classifier reportedly blocked a request to center a div; mentioning animal communication triggered an instant block; a security researcher had a logo-similarity check flagged and bounced to Opus, and even "what is love" got delegated. Users traded a disclaimer phrasing that clears false positives, and worried about long background runs blocked hours in. A recent content policy adjustment drew its own complaints.

Company signals

CFO Krishna Rao described a culture interview no candidate can fail past regardless of raw ability, and said the heaviest internal token users sit in finance, led by the head of tax. Reports put pay above $750,000 for engineers who can build model architectures from scratch. On product, Claude Science aims at lab research and drug discovery. Unverified chatter claimed a confidential IPO filing, while a prediction market opened a line on the US government revoking access to an Anthropic model again and critics revisited Dario Amodei's advocacy for chip export bans.

Google

Google ran in two registers at once. Features kept landing across its consumer surface and people were already stress-testing them, while the thing most talked about — the next Gemini — stayed rumor. Against that, Sundar Pichai gave an unusually blunt answer about whether the AI boom is a bubble.

What the next Gemini might be

Polymarket priced a 62% chance that Google's next-generation Gemini Pro arrives on July 17 (market odds), and bindureddy relayed a claim that Gemini 3.5 already performs somewhere near GPT 5.6 (secondhand assessment). Neither is confirmed by Google.

What is concrete is quieter. Gemini 3.5 Flash turned up in the free tier with a million-token context and no card required (spotted by users), and Google posted tabfm-1.0.0, a tabular foundation model, to Hugging Face (release).

NotebookLM starts making video

NotebookLM's Video Overview generates 60-second vertical clips with voiceover, animation and subtitles from uploaded material, free and with no editing software (feature walkthrough). Usable technique arrived alongside it: one or two strong sources beat ten, and the prompt should direct it like a screenwriter (craft notes), with suggested uses running from single book chapters to plain-language paper explainers (applications). The same tester logged the limits — the hard 60-second cap, no fine script control, generic visuals (shortcomings).

Elsewhere Gemini added one-prompt generation of PDFs, Docs, Sheets and Slides (document output), Omni Flash replaced a video background cleanly in one test (result), and the smart features rolling through Gmail, Chat and Meet drew privacy pushback over what they read (objection).

Agents, small models, a candid CEO

Antigravity moved an entire Vanilla CSS design system to another platform in two to three hours (migration), and Gemini CLI picked up an extension bundling 27 agents and 33 workflows (extension). One developer made the case that sub-70B models are badly underrated, running Gemma 4 at 26B and 31B locally to cut token spend (argument). Asked point-blank whether AI is a bubble, Pichai said yes (the exchange) — landing near commentary that Google's spread across dozens of products is itself the constraint (critique).

xAI

Nothing xAI shipped in this window was a new model. The visible activity was product surface: Grok Imagine widening its reach in video and image work, the Grok Build editor taking an incremental update, and secondhand claims about the company's build speed and its stated design philosophy circulating on X.

Grok Imagine stretches to 15 seconds

The most substantive change was capacity. Grok Imagine can now produce up to 15 seconds of video in a single request, rendered as one continuous take and extendable afterwards. Reactions concentrated on motion quality, with testers calling the dynamic effects strong and passing around a "Return to Mars" demo made with version 1.5. Stills drew attention too: a hands-on upscale of a childhood photo came back clean, one creator ran a Vigeland-inspired concept series, and Grok added templates for the US 250th anniversary timed to July 4.

Build, and the everyday Grok

Grok Build's v0.2.86 release added selectable speech-to-text languages in editor settings and ghost-text autocomplete for the next prompt; a security researcher separately reported strong results on coding tasks. Smaller uses accumulated around the assistant itself, including alt text generated on mention within X and a prompt template for studying kanji. Musk restated the positioning behind all of it, arguing Grok should be maximally truth-seeking rather than evasive. Jensen Huang, meanwhile, repeated his claim that xAI stood up a 100,000-GPU supercomputer in 19 days against an industry norm measured in years.

Microsoft

Nothing Microsoft said publicly during the window mattered as much as a report about what it intends to ship. The smaller items around it pointed the same way: tighten the Copilot surface, and push agent-building material and measurement conventions out to the developer base.

One Copilot instead of two

Microsoft plans to fold its consumer and enterprise Copilot into a single unified application, dropping features that see little use and pulling AI coding into the same product, with August given as the target. The report is attributed to The Information; Microsoft has not confirmed it, so it should be read as a plan rather than a commitment.

Agent plumbing, and how to judge it

Ninety-nine ready-made .NET skills for AI agents went out with almost no announcement, and most .NET developers appear unaware they exist. On measurement, Microsoft's Boris Lebach proposed a unit for token efficiency — speedup per 100M tokens — for comparing inference and serving stacks. Responsible AI product lead Sarah Bird separately argued that trust is context-dependent rather than a fixed property of a system.

Alibaba

Alibaba's day was less about a launch than about how far Qwen has already travelled: adoption research, independent testing, and developers reusing Qwen components inside their own systems.

Open weights and who is downloading them

A paper on open model adoption argues that Chinese releases, led by Qwen, now set the pace: China passed the US in open-source model downloads in the summer of 2025, and by March 2026 the tracked cumulative figure for Chinese models had reached 1.15 billion. Hands-on testing pointed the same way: a role-playing and agentic run across eight local models found Qwen3.6-27B more resilient than peers of its size.

Developers keep treating Qwen as infrastructure rather than a product. One reference-guided image pipeline encodes the reference image through the Qwen VL encoder with the weights left frozen, while a look at how Qwen 3 8B stores days of the week found the first two principal components forming a circle.

Agents, image tools, and safety talk

The head of the Qwen team spent about 25 minutes on where open agent models are heading, covering direction and training method. ModelScope released Civision, an agentic image generation tool that routes tasks from a plain-language description of what the user wants. Separately, Tsinghua's Xu Wei and Fu Hongyu of Alibaba Research Institute discussed frontier safety thresholds and evaluations, including where US-China cooperation might sit.

Mistral

Mistral's window was narrow and technical: one release, aimed squarely at formal mathematics rather than at general assistants, and a first outside attempt to see what it can actually prove.

Leanstral takes on Lean 4

The company put out Leanstral, presented as the first open-source code agent built specifically for Lean 4, the proof assistant used to express complex mathematical reasoning, with automating formal proof work as the stated goal, as introduced by Merve Noyan. Hands-on reaction followed quickly: a Leanstral 1.5 run was pointed at the First Sylow Theorem, and the person testing it noted this was not the hardest thing the model had handled, a result Sophia Yang passed along. That is one third-party observation, not a benchmark, but it is the kind of check that will settle how far the agent goes.