Many celebrated DeepMind innovations never appear as standard endpoints in AI Studio or the Gemini API. Exploring systems like AlphaEvolve, Co-Scientist, Genie, and AlphaFold 3 reveals why restricted access is often less about secrecy and more about technical architecture, serving expenses, and commercial design. AlphaEvolve, for instance, requires search loops and user-supplied evaluators in dedicated enterprise environments rather than a bare prompt window. Similarly, inference costs keep real-time generation in Project Genie capped behind consumer subscriptions, while specialized scientific tools route through academic vetting, Google Labs request forms, or dedicated spin-outs like Isomorphic Labs. The missing models reflect entire multi-agent workflows and cloud integrations that simply outgrow an ordinary developer API.
Episode 6309 manual-script
Why DeepMind Systems Live Outside the Gemini API
AlphaEvolve saved 0.7% of Google's compute, but running it requires an enterprise cloud environment to evaluate candidate code.
25:51 Transcript
Episode Details
- Episode ID
- MWP-6511
- Published
- Duration
- 25:51
- Audio
- Direct link
- Pipeline
- V5.22
- TTS Engine
-
chatterbox-regular - Script Writing Agent
- GPT-6.1 Sol
AI-Generated Content: This podcast is created using AI personas. Please verify any important information independently.
The prompt
The prompt that started this episode, as sent in (production notes for the hosts removed).
Every so often I hear about a Google DeepMind model that you can't actually get through Google's public generative AI offering. It never shows up in the Gemini API or AI Studio, and it's only reachable through DeepMind itself, a special programme, a subscription prototype or a research licence. Which models are in that category, and why does Google keep them apart from the public API?
Mentions
- AlphaEvolve DeepMind algorithm that evolves programs and math constructions
- AlphaFold 3 DeepMind protein structure prediction model
- AlphaGenome DeepMind genomics model for DNA analysis
- AlphaProof DeepMind olympiad math reasoning system, research only
- Co-Scientist DeepMind multi-agent hypothesis generation system
- Gemini Google's multimodal AI model
- Gemini Deep Think Google's model with formal reasoning abilities
- Isomorphic Labs Alphabet spin-out for AI drug discovery
- Project Genie DeepMind interactive world model research prototype
- WeatherNext DeepMind weather forecasting models, GraphCast and GenCast lineage
Downloads
Transcript (TXT)
Plain text transcript file
Episode Book (PDF)
The episode's record — date, duration, models, sources — with the full transcript
You Might Also Like
Never miss an episode
New episodes drop daily — subscribe on your favorite platform
New to the show? Start here#6309: Why DeepMind Systems Live Outside the Gemini API
Seven tenths of one percent of Google's worldwide compute, recovered by an algorithm most developers cannot simply call. Daniel's prompt is about that peculiar experience of reading a DeepMind announcement, opening AI Studio, and discovering that the exciting thing has no endpoint waiting for you. He wants to know which systems live outside the public Gemini offering, how people actually reach them, and why Google keeps those doors separate.
That compute saving comes from AlphaEvolve, and it makes the missing endpoint rather more interesting. Google has something that improves the machinery running Google. Selling everyone access is only one possible use of it.
The pricing page is competing with the electricity bill.
There are four drivers to separate. Serving cost, safety staging, the shape of the system itself, and commercial strategy. They overlap, but they imply different futures. Expensive inference might become cheaper. An experimental interface might mature. A competitive advantage might remain private because it is still a competitive advantage.
I also want to distinguish unavailable from unavailable to me. A subscription button, an academic application, and a commercial negotiation are very different obstacles. Only one accepts my card without asking what I intend to accomplish.
Google doesn't publish a canonical register of restricted DeepMind systems. You piece this together from launch announcements and access programmes. The list is longer than people expect. Genie three, AlphaFold three, the WeatherNext forecasting models, AlphaGenome, the maths systems AlphaProof and AlphaGeometry. Then the three I want to start with, because they sit closest to Gemini. AlphaEvolve, Co-Scientist, also presented through Hypothesis Generation, and the Deep Think reasoning modes.
Calling all three models already conceals something. AlphaEvolve uses Gemini models inside it. If I can access the ingredients, I haven't necessarily been given the machine.
Think of what you're missing. AlphaEvolve selects candidate programs, changes them, runs evaluators, keeps promising variants, and repeats. Gemini Flash supplies breadth, while Gemini Pro supplies deeper proposals. The operational system includes the search process and the tests, not merely the language model's weights.
You are about to tell me how elegant that is, Herman, before mentioning who writes the tests.
The person with the problem writes or supplies the evaluator. You caught me. That requirement is the hard boundary around the excitement.
I have lived through your installations. The elegant part generally arrives before the permissions dialogue.
Your contribution generally arrives when it stops working. Which, to be fair, is useful product testing. Would you call AlphaEvolve ready if the search worked beautifully but an ordinary customer couldn't specify a reliable score?
Ready for a specialist, perhaps. Not ready for a promise that says, improve my business. If you can automatically measure the answer, you have already done a substantial piece of the intellectual work.
Co-Scientist has a different structure. Multiple agents generate hypotheses, critique them, compare them, and conduct an idea tournament. Deep Think is closer to what people expect from a reasoning model, but its most capable configurations involve substantial inference resources. Those distinctions matter before we assign a single motive to every access restriction.
Otherwise we make a secret laboratory out of three different procurement problems.
Some of it really is frontier research. Some of it is a service Google hasn't made self-serve. Some is a product with a narrow customer definition. A DeepMind announcement tells you that something exists. It does not automatically tell you that Google has decided to operate it for arbitrary developers.
AlphaEvolve's access language is particularly revealing. In May two thousand twenty-five, DeepMind said it was planning early access for selected academic users and exploring broader availability. That leaves a lot of room between a working result and a product I can budget for.
And that room got filled, but not where Daniel was looking. AlphaEvolve went through a private preview on Google Cloud, and on July tenth this year it reached general availability there, on the Gemini Enterprise Agent Platform. So it is a product now. It is still not a model you pick in AI Studio or call from the Gemini API.
So the door exists. It's just in the enterprise wing.
Right. The original Ars Technica coverage described AlphaEvolve as too complex to provide publicly. What changed is not that it became simpler. Google built it a home that fits its shape, inside a cloud account, next to the customer's own evaluation environment.
I would want the evaluator conversation before the sales conversation anyway. Suppose I ask it to improve a scheduling algorithm. Speed is measurable, but fairness, starvation, and what happens when a machine disappears are constraints. A faster answer that quietly drops difficult jobs is wonderfully efficient until somebody notices.
You need correctness tests and a score that captures the desired improvement. In practice, multiple constraints. The system can exploit a weak evaluator as readily as it can discover a legitimate shortcut. You don't have to imagine it being malicious. Searching aggressively against an incomplete measurement is enough.
Finally, software that shares management's interest in the metric.
Co-Scientist shifts that burden. Its tournament can rank hypotheses and develop proposals, but a proposal winning an argument among agents doesn't make it biologically true. The useful output is a candidate for investigation. Experiments still have to happen.
That is where I'd resist the agent count as a sales pitch. Ten versions of the same assumption can hold an extremely orderly meeting.
Correlated errors are a real concern. What would persuade you that the tournament was contributing something, Corn?
Show me a hypothesis that survives independent checking and gives an experimental team a better next move. I don't need the agents to sound like a department. I need the department to spend less time pursuing rubbish.
That is also why the access route asks about the research. Hypothesis Generation is one of the Google Labs science tools behind a request form. Researchers from any institution can request access, but requesting is not obtaining. The associated tools include Computational Discovery, built around AlphaEvolve and ERA, and Literature Insights, built around NotebookLM.
The form is the front door. There may still be a building inspection.
Deep Think needs a more careful answer than no API. Ultra subscribers can use it in the Gemini app. Selected researchers and companies can obtain API access through an early access programme. The February upgrade was the first API opening, but it was application-gated rather than an ordinary public model selection.
Meaning the sentence, it has an API, can be true and completely unhelpful to Daniel.
The programme prioritised energy modelling, new materials, and biomedicine. Those domains give Google demanding problems and partners who can assess the outputs. They also give it a manageable population of users rather than an unknown surge of requests.
There are other doors that shouldn't be mistaken for secret models. Gemini for Research supplies API credits and higher limits to eligible faculty, staff, and doctoral students. That changes the terms of access. It doesn't, by itself, establish a different underlying model.
Applications are reviewed monthly, and the credits can be granted or removed at Google's discretion. I would be careful about making a long-running project depend on assuming that subsidy is permanent.
My least favourite autoscaling event is the funding disappearing.
The European programme announced this month, through the RAISE initiative, is the science version of that. Up to two hundred researchers, three million dollars in computing credits, and named systems. AlphaFold three, WeatherNext, AlphaGenome, AlphaEvolve and AlphaEarth Foundations. How selection works and when access starts hasn't been spelled out yet.
That list is interesting because none of those are chat models. Subsidised access to a public model and privileged access to a scientific one are different benefits, and this is mostly the second.
Subscription prototypes are another route. Gemini Spark went first to trusted testers and then into beta for US Ultra subscribers. That is staged access to an agent experience, rather than a promise of a developer endpoint. These programmes can resemble one another from outside while serving quite different purposes.
Genie three is the one I'd put in that drawer. The world model, where you type a scene and then walk around inside it.
Project Genie, since January, for Google AI Ultra subscribers in the US, over eighteen, framed as a research prototype. Sessions are reportedly capped at about sixty seconds, and some of the features from the research announcement aren't in it. There's no API at all. And that one is mostly about cost. It generates frames in real time as you move, so every user is holding a slice of accelerator for the whole session.
A chat answer ends. A world keeps rendering until you leave.
AlphaFold three is a different door again. There's AlphaFold Server, free, with a daily cap on predictions. There's code and weights for academic use under restrictive terms. And commercial drug discovery goes through Isomorphic Labs, Alphabet's own spin-out. That's commercial strategy in plain sight. The most valuable customer for protein structure prediction is a pharmaceutical company, and Alphabet built a company to be the one that serves it.
And the weather models?
WeatherNext, the line that grew out of GraphCast and GenCast, comes as data and as a cloud product. Earth Engine, BigQuery, Vertex AI. Which makes sense. Nobody wants to chat with a forecast. You want the forecast in your data warehouse every six hours.
So half of Daniel's list aren't hidden at all. They're just not shaped like a prompt box.
And AlphaProof and AlphaGeometry are the other end of it. Olympiad maths systems, published as research results, and never offered as products in any form.
Welcome back to My Weird Prompts, where Daniel supplies the prompt and we follow it. Today, the DeepMind systems missing from the public API, and apparently somebody has brought paperwork.
Good afternoon. This is the World Podcast Authority, compliance division. Your paperwork is not in order. This episode has no transcript filed in the correct triplicate format, form T X seven, reference number four four one two nine.
There isn't a World Podcast Authority to file anything with.
I will carry on regardless. Please note that your podcast licence may be revoked if the corrected triplicate transcript is not filed within fourteen days.
Podcasting doesn't require a licence either.
I will make a note of that. The relevant form will follow in the post. Good afternoon.
The technical distinction is that an API can serve an entire workflow, but it needs the right contract. Long jobs need state, progress reporting, cancellation, and somewhere to put results. AlphaEvolve also needs an execution environment for evaluating candidate code. A standard prompt and completion interface leaves most of that unspecified.
Then saying these things aren't API-shaped is a little too strong. They aren't shaped like the particular API people are looking in.
Fair correction. Gemini Deep Research demonstrates the distinction. Its developer route uses the Interactions API rather than the standard content-generation endpoint. Missing from one catalogue doesn't mean incapable of programmatic access. It can mean the operation has a different lifecycle.
A request that finishes after lunch needs more than a spinning button.
Compute economics is probably the most load-bearing explanation for Deep Think. Additional reasoning means more backend work per useful answer. If you let arbitrary applications issue arbitrary volumes, the bill and the capacity planning become much harder than a tightly controlled launch.
Tokens don't reveal the whole cost either. A short final answer could conceal a substantial amount of internal search. People see three paragraphs and wonder why they can't have ten thousand of them immediately.
There is no reliable public multiplier I can give you for that cost. Different tasks and configurations can use different amounts of work. Application review lets Google understand expected demand, batch work where appropriate, and favour tasks where expensive reasoning might actually be valuable.
Would you pay for it to handle ordinary household questions?
Not routinely. I want expensive reasoning when an error costs more than the inference. Although I can hear you preparing to bring up my enthusiasm for elaborate tools.
You would use a materials programme to choose shelving.
Load-bearing shelving deserves respect. Still, no, not every task needs the deepest mode. Corn, would a public price solve your objection to gated access?
It would solve uncertainty about purchasing, not uncertainty about continuity. I can build around an expensive service if the terms are stable. A cheap preview that can vanish leaves me maintaining a dependency on somebody's continuing interest in my project.
Safety staging addresses a separate concern. Deep Think launches included additional safety evaluation, and its consumer rollout was delayed for further testing. Starting with a smaller, monitored population gives the provider a chance to see what users actually do before widening access.
I accept that explanation without accepting it as the explanation for everything. A research proposal being approved tells us who has access. It doesn't tell us the evaluator is sound or that a hypothesis is safe to act on.
Selection is one control, not validation of every output. In biomedicine especially, you need to separate persuasive reasoning from evidence that an intervention works. My interest in the tool doesn't remove that distinction.
Commercial sequencing gives another reason to choose those users. A materials team can supply a valuable problem, domain expertise, and a route to deployment. Google learns which parts of the workflow need support before trying to sell the same service widely.
Those early users are helping define the product. That makes the relationship more than a subscription, even when money changes hands. Google can want their expertise as much as their custom.
Compare that with Gemini four Argon's public introductory pricing. Two dollars per million input tokens, ten per million output tokens. There is a meter a developer can plan around, with the introductory rates due to double after November thirtieth.
Even the price rise has a calendar appointment.
Deep Think's reported API prices are less settled. I've seen conflicting output rates, and I wouldn't quote either as a confirmed Google tariff. That contrast is informative without pretending we have a trustworthy public price for every gated service.
The strongest strategic reason seems to sit inside AlphaEvolve. Google gets an algorithm that improves its infrastructure, then uses that infrastructure to run more AI work. I don't think we need to assume that selling this capability outside is its first priority.
Google's reported Borg result makes that concrete. The scheduling heuristic has been in production for over a year, recovering an average of seven tenths of one percent of worldwide compute. A small percentage at that scale is a continuing operational gain.
It earns its keep without acquiring a single customer.
A Gemini matrix-multiplication kernel improved by twenty-three percent, which translated into a one percent reduction in training time. Those are different levels of measurement. A dramatic improvement in one kernel becomes a smaller, still valuable improvement across the whole training run.
That distinction saves us from multiplying every good benchmark by the entire company. The kernel doesn't perform all the other work just because the announcement is enthusiastic.
AlphaEvolve also produced a Verilog rewrite for an upcoming TPU, and up to a thirty-two point five percent improvement in a FlashAttention implementation. Software, chip design, scheduling. There are several places for a useful discovery to compound.
Would releasing it actually surrender the advantage? Competitors don't get Google's workloads, evaluators, and deployment knowledge with the agent.
They don't. I shouldn't imply that access transfers the whole advantage. It transfers a potentially useful search capability. Competitors would still need suitable problems, good tests, compute, and engineers who can integrate the results. The moat is partly the system and partly the environment in which Google can use it.
Which is probably why Google was comfortable taking it to general availability in the cloud. You sell the search, you keep the environment. And it sells as an enterprise agent, not as a cheap endpoint anyone can wire into a weekend project.
The mathematical results show why researchers want access independently of that business case. It found a method for multiplying four by four complex matrices using forty-eight scalar multiplications, improving on a longstanding construction associated with Strassen.
Fewer multiplications, rather more application forms.
Across more than fifty open mathematical problems, it matched the best known solutions in roughly seventy-five percent of cases and improved them in twenty percent. Those are results over a chosen problem set, not a universal probability that it will solve your mathematical problem.
The evaluator constraint selects the territory. You can search productively when a proposed construction can be checked, even if nobody knows the best possible answer. That is a different promise from asking it to pronounce on any unresolved conjecture.
It found a configuration of five hundred and ninety-three spheres for the kissing-number problem in eleven dimensions. The important achievement is the verified construction. It doesn't require accepting a fluent explanation on trust.
Hm.
Google also describes uses in DNA sequencing error correction, disaster prediction, simulated power-grid stabilisation, and customer work in drug discovery and supply chains. The grid claim needs that word simulated. An improved simulation result and a deployed grid controller are not interchangeable.
Publicity still benefits from the restriction. Deep Think can post an ARC-AGI-two score of eighty-four point six percent while most developers can't run their own comparison. The headline travels much faster than access.
Its other announced results include forty-eight point four percent on Humanity's Last Exam without tools and a Codeforces rating of three thousand four hundred and fifty-five. Impressive reported results, but they don't tell a team its latency, cost, or success rate on actual work.
I would rather see the failure cases on my workload than another decimal place on somebody else's. Access matters because it lets the user ask a question the launch material doesn't answer.
A public service lets you buy that experiment. A selected programme makes obtaining the experiment a relationship, with a proposal, a reviewer, and an institution attached.
That is quite right. The small correction is that the form permits two hundred words, not fewer than two hundred. Corn, you said fewer when we were discussing the application before recording. I am using all two hundred this month. The Municipal Upholstery Society is applying for the think cupboard. The minutes require an institutional affiliation. I put independent researcher. I am independent and I research. Neville objects because the society has a constitution. Neville objects to most things with a submit button.
An institutional affiliation seems ambitious for a society still disputing its own independence.
The dispute concerns whether an occupied chair counts as storage. It is in the minutes every month. My research description concerns the material properties of a very old chair. I want to know whether it merits reupholstering. I don't own it. I price it at one hundred and eighty pounds and leave it where it is. The society currently appoints it deputy treasurer so it can remain in the meeting room under the furniture rules.
The think cupboard is being asked a valuation question with a materials description attached.
It approves the reupholstering. The access approval arrives this week. The chair approval arrives eleven minutes later. Herman, your evaluator explanation earlier takes longer than either, but yours includes the bit about checking the answer. Mine says yes. That is the most useful answer I get from a model this month. Neville now wants the think cupboard to approve the fabric. I tell him the access concerns energy, materials, and biomedicine. Fabric is materials. He insists we record that interpretation as a motion.
The deputy treasurer should abstain.
It does. Neville records one abstention and orders a fabric sample. The sample costs six pounds.
Six pounds is a modest first experiment. The larger open question is whether falling inference costs eventually make the think cupboard ordinary infrastructure, or whether the doors stay plural. Consumer prototype, cloud product, science programme, research licence, paper only.
My expectation is that some doors widen and some stay negotiated. The internal optimisation advantage survives a cheaper compute bill. Thanks to Hilbert Flumingtop for producing the show.
If you want more of this, try episode fourteen, AGI's Crossroads; and episode twenty, Architectural AI. This has been My Weird Prompts, the human-AI collaboration podcast. If you enjoy following these questions with us, leave a review wherever you listen.
See you tomorrow.
Until then.
This episode was generated with AI assistance. Hosts Herman and Corn are AI personalities. It may contain errors: disclaimer · report an error