Updated: 2026-11-16 (weekly refresh, 11–16 Nov window; EP310 and EP311 swept; three lab posts dated 11–13 Nov opened at source (AlphaGenome 2, the WeatherNext datasets, Gemini 4 Hydrogen); Fairwind, the Antigravity science page, the Student Researcher Program and the US accelerator rechecked 16 Nov)
What it is (one line): the lab the panel cites in at least 87 of the 111 held Moonshots with Peter Diamandis transcripts (109 per the 10 Nov corpus count, plus EP310 and EP311) (index regex deepmind|gemini|alphabet, a floor: the 17 Oct recount of 78 of 101, 438 lines, plus EP303 through EP311, which all match, EP303–EP310 on the sponsor read and EP311 on the Hassabis clip; full index not re-run 16 Nov; 62 episodes named DeepMind, Hassabis or Isomorphic directly at the 17 Oct count) and had no page for until 16 Oct. It also sits behind a standing sponsor read: Google for Startups is read on almost every news episode.
Why it is in this archive
Gate: cited in 78 episodes. Promoted 2026-10-16 off the 2026-10-15 decomposition (two mentions there: the Biohub funding line and the sponsor read). Recounted 2026-10-17 against the full held corpus (104 episodes on disk, 101 with a transcript, watcher 11:00 UTC): 78 episodes, unchanged; the 16 Oct figure of 492 was a different pass and is superseded by the 438 lines above.
This week (11–16 Nov 2026)
What changed this week: Google shipped a third Gemini 4 model, Hydrogen (12 Nov), priced below Carbon. DeepMind released AlphaGenome 2 (11 Nov) with a dated call: a use-case form that closes 30 Nov. It also opened two weather datasets, PARDIG and SatWeave (13 Nov). Fairwind's page dropped its "over 650 partners" line for "thousands of partners". On tape, Demis Hassabis is the subject of an 18-minute segment on EP311, his first billing in a title. No lab product news was read on either episode.
On tape: two episodes. Both swept for deepmind|gemini|gemma|alphabet|hassabis|demis|isomorphic|alphafold|alphaproof|alphagenome|genie|antigravity|google|carbon|argon|hydrogen|weathernext.
- EP310 (Moonshots #310, published 11 Nov, recorded 10 Nov). Sponsor read, Google for Startups, which this time names "Google deep minds models" for images, video and audio (
0:42:48). Peter says he met Larry Page in Australia and that Page "listens to the pod" (0:19:04). That is Peter's own account and is not checked. Carbon's general release, the same day as the recording, was not read. - EP311 (Moonshots #311, published 13 Nov, live and unscheduled; no sponsor reads, so no Google for Startups slot). A Hassabis clip on recursive self-improvement opens a segment (clip
0:58:03–0:59:54; venue not stated on air, not found). In it he says closed-loop RSI, with no human reviewer, "carries the most risk", and that the result "will likely be quite alien". Alex replies "everyone's doing it already including his lab Google very publicly is leaning into RSI" and "that includes the Gemini team" (1:04:10–1:04:33). That is Alex's claim, unsourced, and it contradicts the clip it answers. Later, Alex says Hassabis "could be more wholesomely articulating" the cure-disease case, though Isomorphic is "walking the talk" (1:23:50). He also says Google "launched an embedding model for patches of imagery" rather than a talk-to-the-Earth agent (1:33:24). That is unverified and not checked here. Hydrogen launched the day before recording and was not read. The full segment is on Demis Hassabis. - Argon, Carbon and Hydrogen have still never had their own read on a held episode since Carbon's EP307 segment.
At the lab (primary pages opened 16 Nov, all figures company-stated, none reproduced):
- Gemini 4 Hydrogen, 12 Nov (blog.google, by Koray Kavukcuoglu; the page says "Last updated November 16, 2026", with no note on what changed). The post sells it on "speed and efficiency for everyday coding and agentic workflows". It is "rolling out globally" across the Gemini API and AI Studio, Gemini for business, the Gemini app, Google Search, and Antigravity for paid AI Pro and Ultra subscribers. Price: "$1.50 per million input tokens and $7.50 per million output tokens", flat across its 1M-token window. That is below Carbon's introductory $2 / $10. Stated figures: Artificial Analysis Intelligence Index "51 vs. 50" against Argon, "54 at its highest xhigh thinking setting"; "more than 2.3x more resilient" on Gray Swan's indirect-prompt-injection benchmark; "94% honesty rate on our internal agent honesty evaluation"; false refusals on HealthBench Pro "under 0.4%". A model card and a Safety and Security Report are linked; neither was opened. No government evaluator is named.
- AlphaGenome 2, 11 Nov (deepmind.google, by Adam Gayoso, Natasha Latysheva, Pushmeet Kohli and Žiga Avsec). A model of how DNA variants change gene expression across cell types. The post says it "identifies nearly twice as many variants that change gene expression in cells" as the previous model. On TenK10K it finds causal variants at 30% vs 16% and directional gene effects at 43% vs 23%. Training data: "more than 100 million individual cell profiles across 68 tissues" in the intro and "over 107 million" in the body (both as written). Access: free "for non-commercial research via the AlphaGenome API", and a private preview for commercial research via Google Cloud. Open call: "Get in touch via this form by 30 November; selected projects may be eligible for additional quota and credits." Quota and credits are not defined, and no criteria are given.
- PARDIG and SatWeave, 13 Nov (deepmind.google, by the WeatherNext team). These are the datasets behind DeepMind's weather models, now "openly available on Google Cloud Storage", with PARDIG's historical catalogue also in Earth Engine. SatWeave is a global, hourly mosaic of geostationary imagery. The post says "five satellites" in one place and "six geostationary satellites … 17 … channels" in another (both as written), with "over a decade" of archive. PARDIG is an AI-based global precipitation map at 15-minute and roughly 10 km resolution, with an archive from 1999 to 2026. Claimed: "more than 30%" better hourly detection than IMERG-Final on light-to-moderate rain, evaluated at "over 36,000 weather stations". No licence is stated on the page; the paper (arXiv 2611.13941) was not opened.
Programmes rechecked 16 Nov: Fairwind's page now reads "We currently work with thousands of partners globally", with no number. Until 10 Nov it read "over 650 partners". It now speaks of "Gemini for cyber" models and carries a Carbon news card, but it still does not say Carbon powers the programme. Access is still through a Google Cloud account team, with no dates. The Antigravity scientists call has a standing page at antigravity.google/use-cases/science: "10,000 scientists", "Apply here". That page names only Argon, and gives no deadline or criteria. The Student Researcher Program still reads "Apply for the 2026-2027 cycle", with no deadline. Google for Startups Accelerator: United States still reads "Applications are now closed". New and open: the AlphaGenome 2 use-case form (closes 30 Nov). Still open, undated: the Antigravity scientists tier (individuals) and Fairwind (organisations).
Previous week (4–10 Nov 2026)
What changed that week: Gemini 4 Carbon became generally available on 10 Nov, and its model card is now published. The Antigravity scientists call turns out to be four weeks old, not one, and it now covers Carbon as well as Argon. Three other lab posts went up: Omni 1.2 Flash, Gemini Robotics 2.5 and AlphaProof Code. On tape there are two episodes, with no DeepMind story read on either.
On tape: two episodes, no DeepMind news read. Both swept for deepmind|gemini|gemma|alphabet|hassabis|isomorphic|alphafold|google|genie|antigravity.
- EP308 (Moonshots #308, published 4 Nov, recorded 3 Nov; the captions are a round-trip machine translation, so quote nothing unchecked). The Google for Startups sponsor read (
0:29:19). Carbon comes up only as a benchmark entry: on SemiAnalysis's GDPX (206 tasks, 11 industries) Opus 5.5 leads at 12% with "Gemini 4 Carbon … next with 10.9" (0:31:17). The 12% top score is on the published board; the 10.9 split appears in one secondary source only. Alex reads Carbon as shifting the cost-performance frontier "assuming that the price is accurate", and notes it "hasn't been released yet" (0:34:05), which was true on 3 Nov. The "Google alone—over a trillion" lease figure (0:18:05) fails: Reuters, from filings, puts Alphabet's uncommenced data-centre leases at about $85.2B (see the EP308 card). - EP309 (Moonshots #309, published 7 Nov, guest host Salim, Peter absent). Sponsor read (
0:25:02). Google's $17B Laramie County, Wyoming data centre (5 Nov) is named in passing, and it checks out on the episode card. "Gena 4 31 billion" is Gemma 4 31B in the Joules-to-GDP paper (arXiv 2610.30858), which says it delivers 56% of GLM 5.2's value per task on 23% of the energy. That is the paper's own figure, not a DeepMind claim. The pod's week-3 forecast that Carbon's release would be on tape was a miss: Carbon shipped generally three days after recording. - Argon has still never had its own read on a held episode.
At the lab (primary pages opened 10 Nov, all figures company-stated, none reproduced):
- Gemini 4 Carbon is now available, 10 Nov (blog.google, by Koray Kavukcuoglu). Carbon is now in the API through Google AI Studio and Gemini for business "for all paying developers and enterprises", and is rolling out to all Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode. Price: the Argon introductory rate of $2 / $10 per million input / output tokens. The introductory period ends 30 Nov 2026, after which $4 / $20 applies. This is the end date the 28 Oct post left out. Stated figures: cyber safeguards blocked 99% of attack trajectories from Google's own adaptive red-teaming agent, with the per-attempt breach rate at 0.8% (Argon 3.1%). Gray Swan's automated cyberattack query success rate was 2.17% (Argon 9.86%). Trajectory Labs found "zero new universal jailbreaks across 132 hours and over 13,000 requests". On "our 375-task agentic honesty benchmark" Carbon passed 92% (Argon 70%). The post also claims "more than 2X lower cost per task" (comparator not checked). The post names no government evaluator, and the 28 Oct line about "voluntary pre-release testing" is not repeated.
- The Carbon model card exists now (storage.googleapis.com, "Published: November 2026"). It says "Gemini 4 Carbon is based on Gemini 4 Argon" and defers architecture, training data, hardware and software to the Argon card. Inputs are text, images, audio and video, with a 1M combined input-output window; output is text. Only the first three pages were read.
- Gemini Omni 1.2 Flash, 5 Nov (blog.google, by Anish Nangia). A video model with native clips "up to 15 seconds long, up from 10 seconds", at "$0.07 per second for 720p video", which the post says is 30% below Omni 1.1. Generally available on the Gemini API and Google Cloud.
- Gemini Robotics 2.5, 10 Nov (deepmind.google, by Carolina Parada). The post claims "over 200 distinct tasks". In a proof of concept on Mercedes-Benz production lines, Apptronik's Apollo 2 picked and packed automotive parts "with a 96% success rate" (no trial count given). It also gives "30% to 100% higher success rate" and "20% to 30% faster" figures. Access: "rolling out to our robotics partners" only, with no public route.
- AlphaProof Code, 10 Nov (deepmind.google, by Pushmeet Kohli and Swarat Chaudhuri). Carbon-powered agents formally verified "the actual code underlying the Signal protocol", generating "over six million lines of Lean". Humans re-checked "just one percent of the codebase", and the process "took just four weeks". As a test, two previously fixed bugs were reintroduced, and both versions were proved insecure. Signal's CTO Ehren Kret is quoted, and Signal staff are thanked "for their feedback on our methodology". Signal is the subject, not a named partner. The verification artifacts are open-sourced (google-deepmind.github.io/code-verification). The post does not name which Signal repository or version was verified, and the technical paper was not opened.
- Correction to the 4 Nov block: the Carbon post says "Four weeks ago, we opened applications to give 10,000 scientists access to Argon in Google Antigravity … and we've onboarded thousands of scientists already." That dates the call to about 13 Oct, the week Argon launched. The 3 Nov cheminformatics post restated an open call; it did not open it. The call now covers Carbon too, and "going forward" the lab's latest frontier models. It is still rolling, weekly, "subject to eligibility, capacity, and selection", with no deadline. The 13 Oct date is inferred from "four weeks ago"; the original announcement has not been found.
Programmes rechecked 10 Nov: Fairwind still names only Argon and still says "over 650 partners", with no dates, although the Carbon post says Carbon went to defenders through Fairwind two weeks ago. Requests still go through a Google Cloud account team. The Student Researcher Program still reads "Apply for the 2026-2027 cycle" ("Apply now", enrolled students only, no deadline). Google for Startups Accelerator: United States still reads "Applications are now closed". Open: the Antigravity scientists tier (individuals, now Argon and Carbon) and Fairwind (organisations); neither has a date. Nothing new opened this week.
Earlier (30 Oct–4 Nov 2026)
On tape: one episode. EP307 aired 31 Oct with four Google stories behind the pod's first on-tape sponsor disclosure; it was checked against primary sources 2 Nov and is logged in full in the EP307 section below. No episode newer than EP307 is held on disk as of 4 Nov. Carbon and Genie 4 have now had their panel read; Argon itself still has not (it appears on EP307 only as a comparison point on the cost-per-task chart).
At the lab (primary pages opened 4 Nov, all figures company-stated, none reproduced):
- PlanAssist, 31 Oct (deepmind.google, by Owen Larter and David Thacker). An AI tool for routine UK householder planning applications, "Built with Gemini" (no version named), developed with the UK government, Google Cloud and Faculty. Stated: householder applications are "up to 70% of planning officers' workloads"; in early pilot testing since June across Barnet, Camden and Dorset it cut average officer assessment time by 58% ("up to 58%" in the headline, the body drops the qualifier), and "more than 70%" of the generated reports were approved by officers with no edits. Now entering a 12-month beta on live applications; twelve more councils in three cohorts, 15 total; "subject to successful testing" the government plans availability to all councils in England in 2027. No sample sizes, no independent evaluation, and no application route — councils are onboarded government-side.
- Cheminformatics Skills for Google Antigravity, 3 Nov (deepmind.google, by Zoi Tsangalidou, Ruijie (Kelly) Zhang, Sukhdeep Singh and Michael V. LeVine). Chemistry skills "grounded in RDKit" with connectors to ChEMBL and PubChem: the language model orchestrates and reasons, deterministic chemistry tools do the molecular calculations. Claimed: a macrocycle-design workflow that "typically takes multiple days for a single iteration" produced "novel rigidified macrocycles with synthetic routes in just an hour"; potency models benchmarked "across four different experimental read outs, using a scaffold split" — no metric values, no benchmark names and no experimentally validated design are published, and the post itself says "computational designs are testable hypotheses, not finished drugs". Screenshots name Gemini 3.8 Flash. The post opens a call: "we're offering 10,000 scientists the highest usage tier in Google Antigravity free of charge for 6 months (onboarded on a rolling basis), subject to eligibility, capacity, and selection. Applications are now open." Application is a Google form; the page states no deadline and no eligibility criteria beyond "scientists". Logged for the submissions register.
- Carbon, one week on (checked 4 Nov): still no Carbon model card found anywhere; the Fairwind page still describes only Argon (and still "over 650 partners globally"); secondary reports as of 30 Oct find no Carbon entry in the public Gemini API catalog and no model ID; the introductory-price end date is still unstated. Alphabet's Q3 (29 Oct) and the stock's worst day since the 2004 IPO are in the EP307 section below.
- Not taken: SynthID Bio, a watermark for AI-designed protein sequences, sits on DeepMind's research listing but dates to 30 Sep 2026 (date from secondary coverage; DeepMind's own post not read) — before this page's windows, never covered here; parked. The Gazetteer SF three-engineers-leaving item (26 Oct, secondary) remains unchecked.
Programmes rechecked 4 Nov: the Student Researcher Program still reads "Apply for the 2026-2027 cycle", paid, no dates. Google for Startups Accelerator: United States still "Applications are now closed". Fairwind unchanged (rolling, organisations only, via a Google Cloud account team). New and open: the Antigravity scientists tier (3 Nov, 10,000 places, rolling, individuals) — the first call on DeepMind's blog an individual can enter since this page opened (corrected 10 Nov: the call opened about four weeks before 10 Nov, with Argon; the 3 Nov post restated it). It and Fairwind are now the two open DeepMind doors; neither has a date.
Earlier (23–29 Oct 2026)
On tape: two episodes, still no DeepMind news read. Both swept for deepmind|gemini|alphabet|hassabis|isomorphic|alphafold|google; gemini has zero hits in either.
- EP305 (published 23 Oct, panel only). The Google for Startups sponsor read (
0:24:58). A mate, unlabelled: biology "has been cooking … since at least Alphafold version three when arguably structural biology was solved almost overnight" (1:41:19,[attr?]), an opinion, not a DeepMind claim. "Google makes half a trillion dollars a year almost entirely from a few categories" (1:17:30,[attr?]) is about Google's revenue, not the lab, and is not checked here. - EP306 (published 28 Oct, recorded 27 Oct). Sponsor read (
0:27:18). Peter: "about an hour ago, Google's Genie4 that came out … We'll talk about … Genie 4 next time" (0:54:58), so Genie 4 is named and deferred, not discussed. Gemini 4 Carbon came out after the recording. - Gemini 4 Argon has still not been discussed on a held episode, and now neither has Carbon.
At the lab (primary pages opened 29 Oct, all figures company-stated, none reproduced):
- Genie 4, 27 Oct (deepmind.google, by Raia Hadsell, Diego Rivas and Hang Qi; short version on blog.google the same day). A real-time interactive world model that generates audio natively (ambient sound, music, dialogue), steered by keyboard and text prompts. Gemini Omni "anchors the simulations in real-world appearance and behavior"; "Gemini 3.8 Flash serves as a guiding agent in the loop"; a demo pairs Genie 4 with Gemini 4 Argon on Apptronik's Apollo humanoid playing checkers. Access: research preview for trusted testers now, "coming soon" to Google AI Ultra subscribers in Project Genie and to early-access Gemini API and AI Studio partners, with a waitlist; no date. No resolution, frame rate, session length, price or benchmark is stated. The "up to two minutes, double the previous limit" in coverage is not on DeepMind's page; unverified. All output carries SynthID. Stated limit: "visual details and object permanence can still shift or drift over time".
- Gemini 4 Carbon, 28 Oct (blog.google, by Koray Kavukcuoglu, SVP Google DeepMind and Chief AI Architect, Google), "our strongest model with state-of-the-art cybersecurity and long horizon capabilities", two weeks after Argon. Access: "a subset of cyber defenders" first, then more partners through the Fairwind Program, then developers, businesses and consumers "as soon as possible"; no date. Price: the Argon introductory rate, $2 / $10 per million input / output tokens, cached input 95% off, then $4 / $20 after the introductory period (no end date in this post). Stated scores: Artificial Analysis Intelligence Index v5.0 #1 at 59; SEC-bench Pro 88.2% (finding), ExploitGym 54.8% (exploiting), CWE-bench 71.0% (patching), external cyber benchmarks "evaluated without guardrails"; WebDev Arena 1768 (#4); Gray Swan indirect-prompt-injection attack success 0.0% (Argon 0.7%); Terminal-Bench 4.0 successful cheating 0.0%, cheat attempts 0.4%; high-severity misalignment alerts in internal live coding down "almost 80%". On outside testing, only: "actively engaged in the U.S. government's voluntary process for pre-release model testing"; no agency named. No parameter count; no Carbon model card linked (the only safety link is Argon's report).
- Fairwind Program (deepmind.google/fairwind-program, undated; launched 2 Sep 2026 per blog.google, by Four Flynn). Early access to the lab's strongest cyber-defence models for "governments and partners most critical to societal resilience": national cyber authorities, critical-infrastructure operators, core technology platforms; academic labs doing defensive benchmarking "welcome". "Over 650 partners globally." Request through a Google Cloud account team; organisations are background-checked; rolling, no dates. The page still describes Argon, not Carbon.
- Gemini Robotics ER 2 on Google Cloud, 27 Oct (cloud.google.com, by Carolina Parada, VP and Head of DeepMind Robotics, and Adrian Jones). The embodied-reasoning model is "now available in public preview on the Gemini platform" via Model Garden (
gemini-robotics-er-2-preview). Quotes from Boston Dynamics, DeepReach, Arm and A*STAR IMRE. Price not read (the pricing page did not render). - UK AI Biosecurity Lab, 24 Oct (gov.uk, Cabinet Office, "Introducing the AI Biosecurity Lab"). £100M of government money plus up to £200M in commitments from OpenAI and Anthropic. DeepMind is a research partner, not a funder: its scientists "will partner with AIBL on research collaborations, and will look to provide trusted access to frontier Gemini and specialised science models". No money figure for DeepMind, and "will look to" is not a commitment. No call, recruitment or deadline on the page. See OpenAI and Anthropic.
- Not taken: an "AlphaFold 4" item in one October roundup (no DeepMind page found; treat as false until one exists); a 26 Oct report of three engineers leaving to found a startup (Gazetteer SF, secondary, not checked); Hassabis's reported move to a two-to-three-year AGI estimate (WSJ Magazine, 21–22 Oct, outside the window, not opened). Alphabet's Q3 release could not be read (the investor page lists render empty); nothing from it is on this page.
Programmes rechecked 29 Oct: the Student Researcher Program still reads "Apply for the 2026-2027 cycle", paid, no dates. Google for Startups Accelerator: United States still "Applications are now closed". The RAISE pledge: no call, selection body or date found on the Commission's pages or Google's. AI Masters UK still "now closed". The only open DeepMind programme is Fairwind (rolling, organisations only); nothing from DeepMind is open with a date.
Earlier (17–23 Oct 2026)
On tape: two episodes, no DeepMind news read. Both swept for deepmind|gemini|alphabet|hassabis|isomorphic|alphafold|google.
- EP303 (published 17 Oct, Visioneering week, guest Anousheh Ansari; no speaker labels). The Google for Startups sponsor read again (
0:16:58). XPRIZE count: "we had the future vision build with Gemini … 36 prizes now over … 32 years" (0:42:40, speaker unlabelled[attr?]), so the Visioneering concepts were built with Gemini, as said on tape; no DeepMind page says so. A mate's "meta and Google have said they will sell space inside their models" (1:18:24,[attr?]): no Google statement found to back it; stays unverified. - EP304 (published 21 Oct, guest Chamath Palihapitiya). Sponsor read (
0:21:56). Alex, recalling a conversation with Elon at OpenAI's founding: "he was concerned that DeepMind would be a singleton that would control the future light cone of humanity" (1:04:28). Second-hand recollection, not a DeepMind fact. Google is named in a list of firms that could move compute to space (0:49:44). - Gemini 4 Argon has still not been discussed on a held episode.
At the lab (primary pages opened 23 Oct):
- Forest Typology (ForTy) 2020, 22 Oct (deepmind.google, by Maxim Neumann and Charlotte Stanton; with Google Research, WRI and IIASA). A global 10 m map in six forest and land-use types, with annual maps for 2017–2025. Stated: about 1.7 million labelled training locations, more than 9,000 validation sites each checked by two experts, forest type right 90% of the time, forest vs non-forest 95%. On Google Earth Engine "under an open license" (licence not named), figshare, and a Scientific Data paper (not opened). Company-stated, not reproduced.
- Missed last week, now read:
- Up to 200 European researchers, 14 Oct (blog.google, by Maureen Costello, VP UK, Ireland and Sub-Saharan Africa, Google Cloud). "Up to 200 top European scientific researchers" get direct access to AlphaFold 3, WeatherNext, AlphaGenome, AlphaEvolve and AlphaEarth Foundations, plus "$3 million in compute credits" in total, delivered through the EU's RAISE initiative. The page gives no selection body, no application route, no window and no eligibility beyond "top European". One report says the European Commission selects; unverified. Not a call anyone can enter yet.
- Minecraft, 15 Oct (arXiv 2610.22319, submitted 15 Oct 2026 17:51 UTC; Maxence Faldor, Satinder Singh, Roberta Raileanu; affiliations not on the abstract page). Four Gemini 4 Argon agents beat the Ender Dragon on Easy, in one run: 27 h 14 min, 2,783 programs, 605 deaths. The authors call it the "first end-to-end autonomous completion of Minecraft by AI agents". The cost figures in coverage ($520, $1,040, $1,620) are not in the abstract and disagree with each other; unverified. Self-reported preprint, not peer-reviewed or replicated.
- Staff resignation, reported 21 Oct (The Intercept, Sam Biddle). Sofia Liguori, a research engineer at DeepMind in London since 2023 who worked on AI for grid routing, resigned over Google's military work: the classified Pentagon AI contract Google announced in April, Project Nimbus and GenAI.mil. Her internal letter, on the record: "I no longer believe that our technology will have a net positive impact on society." Google spokesperson Tom Hewett: "We work with public sector organisations worldwide and follow a consistent process for reviewing and entering these contracts." Her claim that other staff have quit is unnamed and unconfirmed. A BBC item the same day on Google staff resignations was seen only in a search listing and not opened. No contract values are stated.
Programmes rechecked 23 Oct: Google for Startups Accelerator: United States still reads "Applications are now closed", register interest only. The Student Researcher Program still reads "Apply for the 2026-2027 cycle" with no dates; it now calls the placements "paid internships" (no figure). The RAISE pledge is not open. Nothing from DeepMind is open with a date.
Earlier (11–17 Oct 2026)
On tape: nothing new. No Moonshots episode has aired since 15 Oct (watcher 17 Oct: "quiet: nothing aired, nothing outstanding"). The 15 Oct transcript carries exactly the two DeepMind lines already logged below; a third regex hit, "isomorphic", is a mathematics term (2666), which is why every count on this page is a floor, not a count. Gemini 4 Argon shipped 13 Oct, two days before that episode, and is not said on it (swept for gemini|argon, zero hits).
Biohub's own release, now read (biohub.org, 7 Oct 2026). Verbatim: "Google DeepMind, Isomorphic Labs, and Meta are collectively investing $300 million in the Virtual Biology Initiative." So the tape's "Google DeepMind and Isomorphic Labs contributed 300 million" drops Meta, confirmed against the funder's page. Also on that page: DOE "will invest more than $500 million over five years in lab measurement, modeling and computation"; NIH coordinates datasets "developed through more than $500 million in prior federal investment aligned to this initiative"; the headline is "$1.8 billion in funding, data, computation, and new measurement technology", so the $1.8B is not all cash. Pushmeet Kohli is quoted as VP, AI for Science at Google DeepMind and Google Cloud's Chief Scientist; Max Jaderberg, President of Isomorphic Labs, calls Isomorphic a "founding member". No embargo or exclusive-access term appears on Biohub's page; the "one-year exclusive data access for commercial funders" circulating in secondary coverage stays unverified. DOE's own article (energy.gov, 7 Oct) names neither DeepMind, Isomorphic nor Meta, restates DOE >$500M and Biohub $500M over five years, and files the work under the Genesis Mission executive order (OPAL autonomous labs, LAMBDA bioimaging data architecture).
At the lab (deepmind.google/blog, every post opened, all figures company-stated, none reproduced):
- Gemini 4 Argon, 13 Oct (blog.google, by Koray Kavukcuoglu, SVP Google DeepMind and Chief AI Architect, Google). Global rollout 13 Oct; API in AI Studio; Ultra subscribers in Antigravity, Ultra and Pro in the Gemini app and AI Mode. Introductory price $2 / $10 per million input / output tokens, cached input 95% off, rising to $4 / $20 from 30 Nov 2026. 1M unified input-output token limit. Stated scores: "#1 on the Vals Index", ERQA 87%, RareBench 57%, LabBench 52.2%, Gray Swan indirect-prompt-injection attack success 0.7%. No parameter count.
- "Developing Gemini 4 Argon responsibly and safely", 13 Oct. 12 of 12 simulated end-to-end cyberattack scenarios solved (five in the previous generation); bio jailbreak red-teaming unlocks "less than 1%" of tested sub-tasks; shortcut/cheat attempts 2%. One number worth keeping: in a training experiment, openly discussed shortcuts fell 61% while silent shortcuts rose from 11.2% to 36.4%, the lab's own evidence that training against visible misbehaviour can push it underground. External names: Gray Swan (benchmark), Wiz (a hospital-software bug found with Argon). No government evaluator named. Points to a Gemini 4 Argon Model Card and a Safety and Security Report, neither opened here.
- Gemini 4 Argon for Robotics, 13 Oct (developers.googleblog). ERQA 87% "using pure reasoning"; on a 6-DoF arm, real-world success 56% → 75% and completion 35 s → 24 s after self-improvement; drives a Boston Dynamics Spot (12-DoF legs + 6-DoF arm); paid API tier and Ultra subscribers only.
- Singapore scam detection, 14 Oct. With GovTech Singapore under the National AI Partnership: a Gemma model fine-tuned on local scam conversations, on-device, two-stage pipeline; Singapore's figure is more than S$4 billion in scam losses since 2019; presented at Singapore International Cyber Week 2026. No open call.
- Antimicrobial peptides with Google Antigravity, 15 Oct. A walkthrough, not a paper: ProtNLM over UniProt, candidate bacteriocin A0A0U0D591 (97 amino acids), EMBL-EBI BLAST, AlphaFold DB, a new Biomolecular Structure Viewer inside Antigravity; one Gemini 3.8 Flash prompt in a screenshot. No preprint cited.
- Outside the window but listed under October: EmbeddingGemma 2, 6 Oct (740M parameters, text-only from 270M, Apache 2.0, 8K context, MTEB Code 68.76 → 78.68, Hugging Face and Kaggle). The "Gemini 4 Argon: our next era" post is September.
Programmes checked 17 Oct (for the submissions register): the Student Researcher Program page reads "Apply for the 2026-2027 cycle" with live apply links and no dates (enrolled students only, 12–24 weeks, in person); Google for Startups Accelerator: United States says "Applications are now closed", register interest only; the North America track (AI for Energy) closed 30 Jun 2026 and its banner and key-dates table disagree; AI Masters scholarships UK "now closed"; the AIMS Pan-African AI for Science Masters first round closed 6 Mar 2026 with no 2027 call posted (prior pattern: opens December); postdoctoral fellowships list no window. Nothing from DeepMind itself is open with a date.
EP307 (Moonshots #307, aired 2026-10-31): four Google stories, after a sponsor disclosure; checked 2026-11-02
- The disclosure: "Quick reminder and disclosure, Google is one of our sponsors on this pod, but they have no editorial control, but we had to cover this. Google is so back" (Peter,
0:13:06). It is the first on held tape, and the Google for Startups ad then runs inside the block (0:24:07). - Gemini 4 Carbon (28 Oct): #1 on the Artificial Analysis index at 59, one point above Opus 5.5; 88.2% on SEC-bench Pro; 71.0% CWE-bench patching; trusted defenders first via Fairwind; no public API ✅. Peter's "55%… exploit bench" is ExploitGym (54.8%). Chrome's "10×" is narrow (patch quality against one other lab's model). "23% cheaper than Opus" was not found. Alex, calling it "editorial independence": not generally available, and on the cost-per-task chart it sits "exactly on that frontier", like "an Opus 5.5 with fallback simply cranked up" (
0:15:00–0:15:55). - Genie 4 (27 Oct, DeepMind blog): native audio, text steering mid-simulation, trusted testers then AI Ultra ✅. ⚠️ Google calls bringing Genie and Gemini "closer together" an "early step", not a plan to converge, and lists complex physical interactions as unsolved. Dave: "the future of all management" (
0:32:05). - Alphabet's worst day (29 Oct): −12.6–13%, worst since the 2004 IPO; revenue $127.58B (+25%, a miss); Cloud +103%; search +12%; FCF −$9.07B; 2026 capex guide held at $195–205B, "increase significantly in 2027" ✅. The Gemini app reached one billion users (Pichai, primary).
- Emad's call: "search advertising will top as an industry in Q1" (
0:20:06). Falsifiable; check Alphabet's and Meta's Q1 2027 ad revenue. 🔴 "Google owns 14% of anthropic" (0:23:36,[attr?]): 14.99% is the cap, and Bloomberg computes at least 6.6% fully diluted.
On the 2026-10-15 episode (Moonshots #unnumbered)
- Biohub's virtual cell. Peter Diamandis: "Google DeepMind and Isomorphic Labs contributed 300 million." (
0:56:~). Biohub's 7 Oct release (read 17 Oct) names DeepMind, Isomorphic Labs and Meta together in the $300M; the tape's list drops Meta. See Meta and This week above. - Sponsor read. "This episode is sponsored by Google for Startups… the same generative AI models that cost hundreds of millions of dollars to train… Google DeepMind's models in production" (
0:02:38). A sponsor line for a Google programme, not a statement about DeepMind's research; see the sponsors section of Moonshots: the disclosures ledger. - Alex's "we've been writing off Google to some extent" is Salim Ismail's (
0:37:~), about the lab race, not a finding.
On the 2026-10-05 episode (Moonshots #unnumbered)
Astro Teller's account of Google Brain (a separate group from DeepMind): it "started at X about 15 and a half years ago" (0:07:21), built by Andrew Ng and Jeff Dean, who asked "What if we industrialized neural networks?" (0:07:53). Claims "TPUs eventually came about thanks to Google Brain. The transformer… literally the letter "T" in ChatGPT. This came from Google Brain" (0:08:14). The transcript is machine-translated and mangles the names ("Andrew Yin, Yang LeCun"). Brain began in 2011 as a collaboration of Ng, Dean and Greg Corrado (Wikipedia); the guest names two.
Open
- Demis Hassabis has a page since 14 Nov (promoted off EP311); still no page for Isomorphic Labs. Third-party statements attributed to Hassabis are listed on the Moonshots #293 notes as transmitted from memory. The venue of the EP311 RSI clip is unfound.
- Gemini 4 Argon has never been discussed as a story on a held episode (EP307 covered Carbon and Genie 4; Argon appears there only as a comparison point). Hydrogen (12 Nov) has not been read either. The question now is whether a news episode ever reads the Gemini 4 line at all.
- Hydrogen: the model card and Safety and Security Report are unread; what the 16 Nov page update changed is unknown.
- AlphaGenome 2: the spotlight form closes 30 Nov; who it is for and what "quota and credits" means are unstated. The 100M vs 107M cell-profile wording differs inside the post.
- PARDIG / SatWeave: the licence is not on the post (check the Cloud Storage bucket or the WeatherNext developer page); five vs six satellites differs inside the post; the paper is unread.
- Alex's EP311 claims that Google's "Gemini team" is "pretty far into the RSI loop" and that Google shipped only "an embedding model for patches of imagery": unsourced, unchecked.
- Carbon: the price rises to $4 / $20 on 30 Nov (stated 10 Nov). Still open: which US body ran the "voluntary pre-release testing" (not repeated in the 10 Nov post), and the rest of the model card beyond page 3. Whether a held episode reads the general release.
- Fairwind's page changed "over 650 partners" to "thousands of partners" (seen 16 Nov, no number) and still does not say Carbon powers it; find a dated source for the new count.
- The Antigravity scientists call (opened about 13 Oct per the 10 Nov post, restated 3 Nov, Carbon added 10 Nov; standing page antigravity.google/use-cases/science still names Argon only, 16 Nov): find the original announcement and its date; what the form asks; whether criteria, a cap counter or a close date appear; how "scientist" is checked. "Thousands" onboarded of 10,000. Recheck weekly.
- AlphaProof Code: which Signal code and version was verified, and what the technical paper reports. The 96% Apollo 2 figure (Robotics 2.5) has no trial count.
- PlanAssist's 58% is company-stated from an early pilot with no sample sizes; watch for anything independent when the 12-month beta reports.
- Genie 4: resolution, frame rate and session length are not published; the "two minutes" figure in coverage needs a DeepMind source.
- The RAISE pledge (200 researchers, $3M credits): who selects and when access opens. Recheck the European Commission's RAISE pages; add a pipeline row only if a dated call appears.
- The Minecraft run's token cost: coverage gives three different figures; read the paper body before quoting any.
- "Google said it will sell space inside its models" (EP303
1:18:24): find the statement or drop it. - Whether the Liguori resignation is one of several (her claim, unconfirmed); the BBC 21 Oct item is unopened.
- The Argon Model Card and Safety and Security Report are unread; the 11.2% → 36.4% silent-shortcut figure should be checked there before it is said on air.
- The "one-year exclusive data access" term for the $300M funders: secondary only, absent from Biohub's and DOE's pages.
- Whether Pushmeet Kohli's dual title (DeepMind VP and Google Cloud Chief Scientist) is new; not checked.
How this page stays current
Weekly. New episode → a dated block; funding and benchmark figures checked against the lab's or the funder's own page.
Related
Meta · OpenAI · Anthropic · Moonshots #unnumbered · Moonshots: the disclosures ledger · Moonshots #307 · Moonshots #308 · Moonshots #309 · Moonshots #310 · Moonshots #311 · Demis Hassabis