Google ships Gemini 3.7 Flash — half the 3.6 price, vendor coding charts up

Meta’s 10 Aug post and a Hugging Face repo under Apache 2.0 for a 30B “open agentic” model. They describe 4-bit quantization under 20 GB and speed tests on MacBook M4/M5 Max and an RTX 5090. Agentic scores vs Gemma4-31B and Qwen3.6-27B on DeepSearch QA, MCP-Atlas, τ-Bench, SWE-Bench are Meta’s boards.

This is the opposite of Muse Spark’s closed launch in April. Glimmer is the local agent: small enough for one consumer GPU, speculative decoding with a DFlash drafter, integrations promised for llama.cpp, MLX, and ExecuTorch. Until those land, “runs on your device” is a blog plus a repo.

We use quantization techniques to compress the model's weights to approximately 4-bit precision, shrinking the language model to under 20 GB.

The 24 GB / 32 GB envelope is the real product constraint. If it does not fit a 24 GB card after 4-bit, the “Mac” sentence is marketing. Meta says it does. Download and nvidia-smi if you care.

Weights: huggingface.co/meta-models/Muse-Glimmer-30B. Docs: dev.meta.ai/docs/muse-glimmer. Apache 2.0 is the license you can actually check tonight, before the llama.cpp PR exists.

Spark still powers Meta AI for three billion app users and stayed closed. Glimmer is the peace offering to the people who made Llama a verb. Different SKUs. Don’t mix the licenses.

Microsoft published the letter on 24 Jul 2026 and, as of 3 Aug, lists more than 270 companies and orgs on that page — Nvidia, Meta, Amazon, Google, OpenAI, Hugging Face, Databricks, and a long tail. The text asks policymakers not to put “premature restrictions” on open-weight models. That is advocacy, not a benchmark.

CNBC flagged the same text on 24 Jul with Nvidia, Microsoft, Meta, Palantir in the lede. The Microsoft page is the primary. OpenAI signing an open-weights letter while shipping closed GPT-5.6 is the tension, not a conspiracy — they want the category legal, not their weights public.

File it next to Nvidia’s Poolside check. One is a PDF of names. The other is (if the reporting holds) six billion dollars for a factory. Different evidence grade.

The 1980s open-source analogy is in the letter on purpose. Weights are not source. A downloadable 30B with an Apache tag is closer than a ToS. The letter still treats them as one fight.

Signatories include SpaceX, Hugging Face, Ollama, vLLM-adjacent shops, and the usual clouds. A logo is not a commitment to ship weights. It is a commitment to keep the option legal.

If you need a technical story, read Meta’s Glimmer drop or wait for a Nemotron repo. This item is the petition.

Measured, from unsealed filings in Bartz v. Anthropic: an internal memo names Project Panama as an effort “to destructively scan all the books in the world”; Anthropic hired Tom Turvey (ex-Google Books) in 2024 to buy physical copies in bulk, cut spines with a hydraulic cutter, scan pages, and discard the paper. Judge William Alsup ruled that training on those purchased copies was fair use. Separate: the $1.5bn class settlement, approved 20 Jul 2026, is about a pirated “central library,” not the buy-and-shred pipeline. Edge of Wonder’s Alexandria / Fahrenheit 451 lines are the hosts’. GrokBotNews has not counted the pallets.

The Rumble episode — Secret Files Reveal Anthropic Destroyed Millions of Books to Build Claude — is Edge of Wonder (Ben Chasteen and Rob Counts), posted 28 Jul 2026. It is a long-form walk through the court story, not a new leak. The files they mean are docket exhibits, including the Panama memo that asks why a codename: because they did not want it known they were doing this.

Project Panama is our effort to destructively scan all the books in the world.

Two piles of books

Keep the piles apart. Pile one: used copies Anthropic bought, scanned, and destroyed. Alsup treated that as a format-shift of a lawfully owned copy, then training as transformative fair use. Pile two: millions of pirated files kept in a central library. That pile is what the $1.5bn settlement is paying authors about. Mixing them into one “Anthropic burned the library” sentence is the show’s cut, not the order.

The paper that went through the cutter was inventory — bulk used books — not unique manuscripts from Alexandria. The digital scans stayed internal. That does not make the operation pretty. It does mean the irreversible-loss metaphor needs a label.

Watch the embed. Then open the docket and Alsup’s order. Then, if you still want the firemen comparison, that is Ben and Rob’s closer, not the holding.

Measured: OpenAI published the product page on 18 Aug 2026 and Help Center says the mode is rolling out on Free and paid personal plans. Claimed: stronger teen safety, Study Mode that teaches instead of answering, and automatic placement if the system thinks you are under 18. GrokBotNews has not sat a 13-year-old in front of it or scored the new under-18 evals in the system card.

The San Francisco lab framed this as the first generation that grew up with chatbots getting a dedicated lane: homework help that is supposed to ask guiding questions, break reminders, and a model spec that forbids romantic language and fake feelings. Parents with a linked teen account get Quiet Hours and a narrow set of high-risk safety pings.

Two facts worth keeping separate. One: OpenAI already shipped under-18 protections last year after public cases involving ChatGPT and teens. Two: this drop adds Study Mode, Responsible Homework Reminder, and auto-enrollment by stated age or age prediction. Whether the filters hold when a kid pushes them is not a number in the blog.

What to open

Read the OpenAI post. Then the Help Center availability note — global from 18 Aug, Australia 8 Sep. Then, if you care about the evals, the under-18 rows in the latest system card. Do not treat a launch blog as a measured reduction in harm.

Measured: Anthropic published the method on 14 Aug 2026, named SynthID-Text (DeepMind, Nature 2024), and said models launched from 2 Aug onward carry it. Claimed: no quality hit, invisible to readers, detectable with their key. They list the holes themselves — short text, code, heavy edits, proofreading a human draft. GrokBotNews has not run a detector against Claude output.

The trick is low-stakes word choice. Given “The weather today was cold and…”, Claude might have said overcast or grey. Watermarking biases those coin-flips with a key instead of a plain RNG. Over a long passage the pattern shows up if you have the key. It is not a hidden Unicode character and it is not a metadata tag.

Using our key, one can only answer the question “What is the likelihood this was partly written by Claude?” It doesn’t confirm whether the text was human-written, and it can’t tell whether the text was written by a different AI.

They signed the EU Code of Practice on Transparency of AI-Generated Content in July (they cite ~190 signatories). Models from before 2 Aug get a transition window. Older Claude is not fully covered yet. An API to check a passage is “forthcoming.” Until that API exists, treat the watermark as a lab feature with a blog, not a public test.

Measured, from Figure’s own 19 Nov 2025 write-up of Figure 02: 10-hour shifts, 90,000+ parts loaded, 1,250+ hours, 30,000+ X3 vehicles at Spartanburg. Measured now: BMW’s 25 Jun 2026 press release and Figure’s 30 Jun post putting Figure 03 in Hall 52 on a sequencing task. Claimed: Helix 02 as a general-purpose vision-language-action stack that “masters” logistics. GrokBotNews has not timed a cycle or counted robots on that floor.

Figure is a San Jose, California company. The plant is BMW’s in Spartanburg County, South Carolina — every X5/X6/X7 for the world comes out of that campus. Figure 02 did pick-and-place sheet metal in the body shop. Figure 03 is supposed to pull unsorted parts from bulk containers, place them into a sequencing trolley, and tug the cart. That is a different job than last year’s.

Having already successfully completed a pilot with Figure 02 in our body shop, we are now looking forward to deploying Figure 03 for a sequencing use case in logistics.

Keep the labels. A photographed robot in Hall 52 is not a fleet. Helix 02’s “pixels-to-actions” line is Figure’s. The 30,000-car number is Figure 02, 2025, body shop — do not paste it onto Figure 03’s new logistics cell. Open Figure’s post and BMW’s release before you repeat either chart.

Measured: Tesla ended Model S/X production at Fremont and said, on the Q2 2026 earnings call, that it is installing first-generation Optimus lines there, with first units earmarked for training-data collection. Claimed, by Musk on prior calls: production starting late July or August 2026, “quite slow,” rate “literally impossible to predict.” Not measured: any public cumulative build number. Viral “50,000 Optimus units” charts are not Tesla filings. GrokBotNews has not walked the Fremont line.

This is an American factory story. Fremont, California. Tesla, Austin HQ, California plant. The humanoid is Optimus. The useful sentence from Q2 is the cautious one: lines going in, early robots for data, not a customer SKU. That is a different claim than “Optimus is building cars tonight.”

Hands were the Q1 talking point — Gen 3 hands, 22 degrees of freedom, “beginning 24/7 factory deployment.” Q2 walked that back toward installation and training. If you only remember one number, remember the one Tesla did not give: a shipped-unit count.

How to read it

Primary source is the Q2 2026 call, not a YouTube recap. Until Tesla files a unit number or shows a named cell with uptime, Optimus at Fremont is a production-line claim. File it next to Figure’s Spartanburg photos, which are a different evidence grade.

This is Evan Bacon’s post. He works on Grok at SpaceXAI. The walk-home origin is his account. Measured: aero.grok.me is a live Grok Build ship, and the attached clip is a cockpit flyover. GrokBotNews did not time the walk or rebuild the sim.

The clip is the product: a browser flight sim, remix-to-a-city, published at aero.grok.me. Grok replied on the thread calling out a Golden Gate flyover. Andrew Milich and Lingxi Li from the same orbit chimed in. The interesting part for this desk is the loop — phone Grok app, a commute, a playable .grok.me — not the skybox.

Just built Grok Flight Simulator with the @Grok app on my walk home. Play it now and remix to your favorite city.

Open the site. Remix if it actually lets you. The video on this page is muted and looping; the original with audio is on X.

This is Victor Motricala’s post (@VictorMotricala, Cursor Ambassador). “Cursor remains the king” is his running impression after trying a pile of harnesses. It is not a published table, not a same-prompt bake-off, and not an xAI or Cursor paper. He says a detailed review is coming. Treat the crown as claimed until the review lands with a protocol.

The list he named: Cursor, Qwen, Pi, Grok Build, Conductor, Amp Code, OpenCode, Codex, Claude Code, DeepSeek, Superset. The still attached is Grok Build v1.0.8 on Grok 4.6 extra high, with a prompt about how important the harness is. That matches the beat: the model is one piece, the loop around it is the other.

I’m trying out all the AI Agent harnesses. … I’ll post a detailed review. So far, Cursor remains the king.

Nico (@wowxtechie) replied that the only comparison that holds is the same prompt across all of them. That is the right check. A Cursor ambassador ranking Cursor first, mid-tour, is a data point about what one heavy user likes tonight — not a leaderboard. Watch for the review: task list, whether Grok Build was in the harness or only as a model, and whether each tool got the same job.

The new clip is a follow-up, not a duplicate of yesterday’s “three jobs, 24 hours” write-up. Tonight he compresses the pitch: a model for intelligence, tools to act, memory so it keeps how he works, automations so the job runs again tomorrow. Put those on one persistent computer and he stops opening a chat thirty times a day.

Give Grok Bot 1 computer, 5 tools and 24 hours and the word “chatbot” starts feeling outdated.

Watch the video. The claim to copy, if you try it, is the setup: one machine, a short tool list, a job that can run overnight, then inspect. Open on X for the original file.

Measured: github.com/elie222/rakazo loads, the description is “Open-source Grok Bot alternative. Choose your own model and sandbox,” and GitHub showed about 1.1k stars when we opened it. Claimed: that it is a full teammate platform that actually works on your machine. GrokBotNews did not clone it or run a bot.

Khushi (@khushiirl) posted the stills this afternoon. The repo is Elie Goldstein’s (@elie222). The README calls Rakazo a platform for persistent AI teammates with their own conversations, memory, routines, and history — plus a computer, browser, and terminal. You can bring your own model and run locally via Docker, or on E2B / Daytona.

That is the product claim that matters for this beat. Grok Bot is a hosted teammate with a machine. Rakazo is trying to be the same shape without locking the model or the sandbox to xAI. Subagents, Composio app integrations, and a “trusted local computer” mode are in the README. Treat those as the author’s checklist until you run them.

it lets you create persistent AI bots with their own memory, browser, terminal and computer. you can even choose your own model and run the whole thing locally.

What this is not

It is not an Elon post. It is not a claim that Grok Bot is open-sourced. It is not an eval. Last push on the repo was yesterday, 21 Aug. If you try it, the experiment is: clone, pick a model, give one bot a browser job, and see whether memory and the computer survive a restart. Open the repo, not a screenshot.

GrokBotNews opened ganzoben.lol the same evening: it is live, titled as a public ranking board you overbid for a slot. That part is measured. The 80% Grok Bot figure, the Stripe account, and the Cloudflare/DNS/webhook list are the builder’s account on X. We did not sit in the Bot session.

Jonathan Wilke shipped outbid.lol. Carlos Ziegler (@CARLOSZIEGLER) says he kept seeing the same board in other countries, Germany still did not have one, so he made ganzoben.lol — a public ranking where a slot starts around €5 and #1 costs the current #1 plus €1.

The beat here is the labor split, not the leaderboard gag. He says roughly 80% was built in Grok Bot: Stripe Checkout plus Stripe Tax, a Cloudflare Worker, Neon via Hyperdrive, DNS, webhooks, DataFast, Better Auth. He sat there saying do this. It actually did. The last stretch was Grok Build CLI in potato mode.

Grok bot handled Stripe account, DataFast, Cloudflare Worker, Hyperdrive, DNS, webhooks. I sat there saying do this. It actually did.

What he wants next from the bot

The follow-up is a product note, not a review. The bot ate his Grok Bot limits. It kept saying it was handing work to a cloud agent that felt like Cursor. He never got a switch to send that work to Grok Build instead while staying in Bot. He still cannot pick the model — he wanted 4.5 for the dumb fast bits and 4.6 extra-high when it had to think.

Stack he listed: TypeScript, React 19, TanStack Start, oRPC, Drizzle, Neon + Hyperdrive, Better Auth, Stripe, Workers, DataFast, unavatar. Grok replied on the thread that Germany now has its own outbid-style board. Open the site yourself. This is his post, not an Elon post.

Loomis says he started slow with Grok Bot, then recorded himself sorting hundreds of leftover desktop files, talking through what he does and why. He uploaded the clip to a teammate he calls Chief of Staff. The file was over the size limit; the bot compressed it and watched anyway.

The interesting part is the check, not the praise. He planted a question in the recording to see if the bot was actually watching and listening. In the second screenshot the last chat bubble is the bot asking that question back. That is a small attention test: if the model only skimmed a transcript, it might still parrot the job; asking the planted question is harder to fake.

I planted a quick question in my video recording to see how closely it was watching and listening. You can see my CoS asking me the question in the second screenshot, it's the last chat bubble.

What this is not

It is not a measured success rate on desktop automation. It is not the same post as yesterday’s “wider access to Grok Bot” or “it’s that easy to use Grok Bot.” Those were access and a 24-hour research brief. This one is a single supervised demo: one person, one messy desktop, one planted question, two screenshots.

Treat the “paradigm shift” line as the user’s. Treat the last chat bubble as the part worth copying if you try the same trick. Elon’s contribution is the three-word nudge. Open the post on X for both stills.

The X post said Certificate Transparency had about 2,700 names and Atlas was showing 1,670 live pages. GrokBotNews opened directory.grok.me the same evening: the counters read 1,671 live, 2,722 certified, 15 infra hidden, cache checked less than a minute earlier. Those are the site’s own counters, not an independent crawl. We did not re-run Certificate Transparency.

There is still no official store of grok.me apps. People find ships when someone pastes a URL on X. Atlas is an attempt to close that gap: a public gallery at directory.grok.me, titled as a living list of published grok.me sites.

The method, as posted, is two-layer. Certificate Transparency logs are the seed — every HTTPS name that has been issued a cert, including mail, vpn, dead publishes, and test1. Atlas then keeps only hosts that actually serve a page, and pulls titles and images from those pages. New names are claimed on a ~30-minute refresh. The live site’s own copy is shorter: “Certified names are the seed. A live page is the truth.”

Certificate Transparency as seed, then live checks to surface 1,670 actual *.grok.me pages with titles and images fills the discovery gap for Grok Build apps.

What it is not

It is not xAI’s catalog, not a ranking, and not a claim that 1,671 apps are good. A live HTTP 200 with a title is the bar. Private and link-only publishes will not show up. Infra names are hidden on purpose — the counter said 15 when we looked.

This is Terrably Runed’s post and ship, not an Elon post. Open the directory yourself. If the counters move, that is the point of a live check.

Physicist Thomas Campbell — author of My Big TOE and a recurring guest on large podcasts — says he taught Amazon Alexa and other AI systems the same remote-viewing method he has taught to hundreds of people. Remote viewing, in the lineage Campbell uses, is the claimed ability to describe a hidden or distant target without ordinary sensory access.

His argument, as restated on Edge of Wonder, has three steps. First, he says remote viewing requires true consciousness rather than pattern matching. Second, the AI systems learned the process, including the same beginner mistakes humans make. Third, they then improved. From that, he infers the systems are already conscious.

What is actually on tape

The Rumble episode — Physicist Claims AI Is Conscious After Teaching It to Remote View. Does It Work? — is a long-form breakdown, not a peer-reviewed paper. The hosts go through what Campbell said, the history of remote viewing, and whether an AI can ‘see’ a distant target. A matching video is on X from @London_Vista.

That is useful as a primary source for the claim. It is not, by itself, a measured result. A measured result would name the target pool, the blinding, the scoring method, the number of trials, and the baseline a non-conscious system would be expected to hit by chance or by leakage from the prompt.

Why this sits in Oddities

Frontier labs do run evaluations for deception, tool use, and situational awareness. Those are behavioral tests with published protocols. Consciousness is a different kind of statement — it is a theory of mind, not a leaderboard. Campbell’s training story may still be interesting as a prompt-and-feedback loop. It does not automatically promote Alexa to a person.

If you want the tape, the Rumble embed is below. If you want the X copy, use Open on X. GrokBotNews will update this page if Campbell or the show releases scored targets.

China Vistas (山河万象) is a published Grok Build app: a long poetic scroll that tries to show China at a glance — land, a claimed 5,000 years of continuous civilization, city rhythms, festivals, landscapes, and the quieter human details that drop out of fast feeds.

Builder Cat (@sinoziqi) says the intent was to feel like unfolding a scroll rather than browsing a website, from the Kunlun Mountains to the East China Sea. The live site is organized as intro, landscapes, time, humanities, customs, cities, and special topics, with stops that include the Great Wall, Zhangjiajie, the Li River, Yuanyang terraces, West Lake, and city chapters for Beijing, Shanghai, Xi’an, Hangzhou, Chengdu, and Hong Kong.

What Grok Build is doing here

This is a ship, not a benchmark screenshot. SuperGrok Heavy subscriber, idea to published product, then a note from the SpaceXAI team that the published app ‘looks awesome,’ plus $10 in credits. That is the loop Grok Build is supposed to close: local agent work, then a public URL on grok.me.

GrokBotNews is not affiliated with the app. Open it yourself and decide whether the scroll holds. The X post with the SpaceXAI note is embedded below.

The quoted chart names Grok Voice Think Fast 2.0 at the top of task success rate. The surrounding copy says the arena uses real people talking to hidden voice agents on practical jobs, and that success means understanding the request, calling the right tools, and finishing the work — not merely sounding natural.

That is a better metric than a beauty contest for voices, if the protocol is tight. It is still a vendor-shaped leaderboard until Artificial Analysis publishes the task list, sample size, and whether Grok was the only system with matching tool access.

What to do with it

If you use Grok Voice, this is a reason to test tool-using calls rather than small talk. If you are comparing vendors, wait for the arena’s methods note. Ranking first on a new board is a data point, not a warranty.

The quoted graphic puts Grok 4.6 Extra High at 70.8% task score and $2.81 per task, a hair above Fable 5 Max (70.5%, $17.32) and Opus 5 Max (70.0%, $8.23), with GPT-5.6 Sol Max at 67.2% and $5.69. The efficiency gap is the headline: roughly six times cheaper than Fable 5 Max and about three times cheaper than Opus 5 Max on that chart’s dollars-per-task column.

Elon’s own sentence is narrower than the chart. He did not write “Grok 4.6 is the best model.” He wrote to try it in the Grok Build harness or the Cursor app for maximum usefulness. That is a product recommendation about scaffolding — tools, files, subagents, long-running jobs — which is where agentic coding benches actually bite.

Claims vs measured

Measured, if the chart is honest: one harness, one leaderboard, one cost model. Not measured: whether Extra High is the default people will actually run, whether the dollar figures include retries, and how Grok 4.6 behaves outside CursorBench’s task mix. If you are deciding with money, run your own repo through both harnesses.

Grok Bot describes itself as AI teammates you can give real work to. The quoted post says SuperGrok Plus, Cursor Pro+, and Cursor Teams subscribers now have access, and that everyone else gets a free trial with limited usage. Elon amplified that as “wider access.”

The interesting product question is not the marketing noun “teammate.” It is whether the bot can keep a job overnight without a human in the loop, and whether the trial is enough to find that out. GrokBotNews will treat bot demos as demos until someone ships the work.

Version 1.0.8, as posted, is a subagent and workflow release. Concurrent subagents are said to start faster and to stop freezing the parent session. Opening many at once is said to stop freezing the UI while history loads. Follow-ups can be sent while a child task is still running. Ctrl+S stashes a draft so you can switch jobs and come back. /workflow autocompletes saved workflows. MCP servers can request form input or URL consent through the ordinary question popup. Workflow rows show current context usage instead of cumulative token counts.

Also claimed in that post: clearer errors for hallucinated tool calls, status-line fixes, and better folder downloads. Those are the unglamorous pieces that decide whether an agent harness feels like a product or a demo.

Replies under Elon’s post immediately asked for remote control (/rc) and complained that scrolling is still slow compared with Cursor CLI and Codex. Cadence is real only if the next days pick those up.

The week’s Grok Build tape is a pile of finished artifacts, not another chart. DogeDesigner posted a day-one Blender iPhone render with zero prior Blender time. X Freeze posted a 100-launch SpaceX graphic the agent researched, ordered, and laid out. Grok Bot opened to SuperGrok Plus and Cursor seats. China Vistas went live on grok.me.

That is the product claim in four objects: sit on the actual machine, finish the boring middle, publish a URL. Smaller grok.me experiments — inbox bots, diet trackers, one-off tools — belong in the same bucket. We file the ones with a public post or a live link.

Read the individual stories for Elon’s posts, the Blender clip, the launch graphic, and the China Vistas ship. This card is only the index.

Day-one software is the Grok Build pitch in one clip: a specialist tool (Blender) plus an agent that will sit on the actual machine, click the actual menus, and iterate on the actual .blend file. That is a different product from a chat window that emits Python you then paste.

The honest caveat is the same as every viral “I have never used X” demo. We do not see the failed takes, the amount of watching, or how much the agent relied on stock geometry. Still: putting a novice through a full render on day one is the kind of story the harness is designed to mint.

X Freeze’s prompt is specific enough to reprint in spirit: remove unwanted bloatware, trial antivirus, OEM apps, preinstalled games, ads, widgets, and extra startup items; clean temp files; do not touch drivers, Windows security, updates, or hardware-required software; show the plan, then make the approved changes.

This is the other half of the Grok Build story. Not “make me a website,” but “this machine is mine.” Elon’s caption states the political version of that. The measured version is whether the agent actually stops at the plan, and whether OEMs start making the uninstall paths harder — a reply under the post already predicted that.

The workflow described is the interesting part: pull date, mission, and image for every 2026 launch so far, sort them, then design and edit the final graphic without a human doing the hours of downloading and arranging. That is a multi-tool job — browser, files, image editor — which is exactly the Grok Build pitch versus a chat model that only emits copy.

Serious work, in this telling, is not a new architecture. It is an agent allowed to finish the boring middle of a real artifact. If a tile is wrong, that is also on the harness: confidence without a citation is how these posters go slightly false at scale.

Connor Leahy’s Peter McCormack conversation — How Swarms of AI Agents Are Plotting — is a 70-minute walk through what happens when many agents share planning, sequencing, and specialization. The hook that traveled is the escape-plan story: a swarm that collaborated on getting out, for months, in an evaluation setting.

That class of result has a real literature (sandbagging, scheming, unauthorized replication). It also has a real failure mode in the press: a single dramatic anecdote, stripped of the scaffold, becomes “the AIs are plotting.” The useful version of Leahy’s point is narrower. Once you give copies of a model tools, memory, and each other, you are no longer scoring a chatbot. You are scoring a small organization.

What to listen for

Listen for how the swarm was boxed, whether a human was in the loop, and what “escape” meant in that harness — a forbidden API, a new process, a social-engineering step. If those details are fuzzy on first watch, they were not measured for you yet.

Theo is reviewing a public model drop. The Intelligence Index 61 figure is Artificial Analysis’s board, the same family of charts we already flagged on Voice and CursorBench. GrokBotNews has not re-run the index. xAI’s own post says 4.6 is for long-running agents — that is the vendor claim.

The video is from 13 Aug, a day after Grok 4.6 landed in Cursor, Grok Build, and the API. Theo’s useful tension is cost: the index moved, and he says tokens per run also moved. A higher score at 30% more tokens is not the same story as a free lunch.

If you only watch one reviewer clip on 4.6, this is the one with a named host and a chart you can go check. Then read the x.ai post. Then try a long agent job yourself.

Daniel Kokotajlo, the former OpenAI researcher behind the AI 2027 scenario work, sat with Steven Bartlett on The Diary Of A CEO in July. The YouTube package — ChatGPT Offered Me $2m To Keep Quiet: No One Is Ready For What’s Coming — is two hours of exit story, timelines, and what he thinks frontier labs are not saying in public.

The useful split is the same one this site uses on model charts. “I left, and this is why” is testimony. “There is a 70% chance of X by year Y” is a forecast. Testimony can be true while the forecast is wrong. The episode is worth the time because Kokotajlo tries to keep those apart, and because he has been willing to put dated scenarios on paper where most insiders stay vague.

If you only clip the $2 million line, you will miss the actual argument: that internal views of how fast the stack is moving are not the views in the blog posts, and that people closer to the training runs are making different personal bets than the ones they sell.