Episode 6308 custom

How Google DeepMind Absorbed the Gemini App

In October 2024, Google moved Sissie Hsiao's Gemini app team directly into DeepMind, turning a research lab into a consumer product division.

Oct 27 26:32 Transcript

With

0:00

local-ai ai-evaluation

Episode Details

Episode ID: MWP-6510
Published: Oct 27
Duration: 26:32
Audio: Direct link
Pipeline: V5.22
TTS Engine: chatterbox-regular
Script Writing Agent: GPT-6.1 Sol
Topics: local-ai ai-evaluation

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).

I've never really understood what Google DeepMind actually is as an organisation. How is it separate from Google generally — is it a division, a subsidiary under Alphabet, a research lab, a product group? How big is it compared with the rest of Google, in people and in influence? And does Gemini actually live inside DeepMind, or is that built somewhere else and DeepMind just does the research? I'd like a clear picture of where DeepMind sits and how it got there.

Google DeepMind has evolved far beyond a secluded British research outpost into Google's unified engine for artificial intelligence. While DeepMind Technologies Limited remains an incorporated UK subsidiary with independent statutory filings, the operational organization reflects deep integration. Following the pivotal 2023 merger of Google Brain and DeepMind, the unit assumed broad responsibility across foundational models, including Gemini, Gemma, Imagen, and Veo. A major organizational milestone occurred in October 2024 when the Gemini consumer application team officially moved inside DeepMind, closing the distance between model developers and product managers. With Demis Hassabis serving as chair and Alphabet's Chief Scientist, operational leadership sits with Koray Kavukcuoglu, who reports directly to Sundar Pichai across research, development, and user-facing apps.

Sources

What the research for this episode read before the script was written. Primary sources first.

Context

Adjacent ground already covered: episode #6037 "Gemini 4 Argon: Google's Million-Token Flagship" profiled the model itself, and #1067 "The 3,000-Person Army" covered how many people it takes to ship a frontier model. Keep both to a brief callback at most — this episode is about DeepMind's organisational structure and its relationship to Google/Alphabet (history of the 2014 acquisition, the 2023 merger with Google Brain, reporting lines, headcount, where Gemini is built and where the product teams sit), not about Gemini's capabilities.

Mentions

AI Mode Gemini-powered conversational Search feature
Discovery Loop New venture by Jeff Dean and Sanjay Ghemawat
Gemini Google's multimodal AI model
Gemma Google's open-weight model family
Google Brain Google research team merged into DeepMind
Google DeepMind AI research lab behind Gemini
Imagen Google DeepMind image generation model
Isomorphic Labs AI drug discovery company led by Hassabis
Lyria Google DeepMind music generation model
Veo Google DeepMind video generation model

Downloads

Episode Audio

Download the full episode as an MP3 file

Download MP3

Transcript (TXT)

Plain text transcript file

Episode Book (PDF)

The episode's record — date, duration, models, sources — with the full transcript

#6308: How Google DeepMind Absorbed the Gemini App

DeepMind usually gets described as Google's research lab, which makes the Gemini app sound as though it arrived from another department. That picture is out of date. Daniel wants to know what Google DeepMind actually is, where it sits inside Google and Alphabet, and whether Gemini belongs there or merely borrows its research. The confusion makes sense. The company name, the legal company, and the organisation doing the work don't map neatly onto one another.

The cleanest starting point is that DeepMind is a real British company with its own accounts, while Google DeepMind functions as Google's central AI organisation. Both descriptions are useful. Neither gives you the whole reporting structure.

A subsidiary with division responsibilities and research lab stationery.

Plus a consumer product organisation. That last part matters enormously. Building the model and running the application used to be more visibly separate responsibilities. Now the people developing Gemini, the frontier researchers, and the Gemini app and developer teams sit under the same leadership.

You can see why Daniel wants an actual picture rather than another logo. Google keeps the familiar names while moving the walls.

August's leadership changes moved a particularly important wall. Demis Hassabis stepped away from day-to-day leadership, and his successor reports directly to Sundar Pichai.

Before we decide whether that's elevation or absorption, the word subsidiary needs some care. People hear it and imagine a company that could pack its bags and leave.

Ownership doesn't provide that freedom. DeepMind Technologies Limited is incorporated in the United Kingdom, ultimately owned by Alphabet. There's a holding-company layer, DeepMind Holdings Limited, in the ownership chain. Describing the legal company as simply a department of Google LLC loses that distinction. Describing its working organisation as independent loses the other half.

The legal company can sign things, employ people, file accounts. The operating organisation decides who spends the compute budget and who has to answer when Gemini misses a release. Those responsibilities can cross legal boundaries.

They can, and in large groups they routinely do. Google DeepMind also names an organisation assembled from teams with different histories. You shouldn't assume every person working under that banner has exactly the same legal employer. I don't have an employee-by-employee map, and a public brand wouldn't establish one anyway.

You're about to draw nested boxes, Herman. I can hear you reaching for the holding company.

I was reaching for the difference between ownership and management.

Your favourite kind of box.

You asked for an actual picture. I am being unfairly punished for supplying edges.

The management change was announced on August fifth. Hassabis became chair of Google DeepMind and Alphabet's first Chief Scientist. Koray Kavukcuoglu became senior vice president of Google DeepMind, reporting to Pichai. Hassabis also remains chief executive of Isomorphic Labs. That is broader scientific scope for him, alongside less direct operational control of DeepMind.

Calling it a demotion would flatten several different jobs into a scoreboard. A chair, a chief scientist, and an operating head hold different kinds of power. I'd rather know who can change a launch date.

That's why the organisation matters outside corporate housekeeping. Who can allocate resources, set a product deadline, accept a deployment, or stop work? Those answers affect the technology people receive. A retained company registration doesn't answer them.

DeepMind's original registration takes us back to September twenty-third, twenty ten. Hassabis, Shane Legg, and Mustafa Suleyman founded it. That's before Alphabet existed, which makes the modern shorthand slightly awkward. Google bought it. Alphabet eventually became the ultimate parent.

Early backing included Peter Thiel, Elon Musk, and Jaan Tallinn. Google acquired DeepMind in January twenty fourteen. The reported price ranges from four hundred million to six hundred and fifty million dollars. Even that spread is a reminder that a transaction can be historically important without every detail becoming public.

Quite a range for something already purchased.

The strategic attraction was a concentrated research organisation pursuing general-purpose intelligence. DeepMind brought its own leadership, research culture, and London base. Keeping a separate legal entity supported that identity. It didn't create an unbreakable promise that the parent would never integrate its work.

I'd put more weight on the leadership than the registration. A legal shell preserves the possibility of a boundary. People with authority have to defend it. Once everybody's budget and promotion depend on the parent's priorities, the boundary can remain on paper and disappear from daily life.

What would persuade you that the boundary still mattered, Corn? Suppose the subsidiary has its own accounts and offices, but the parent appoints its leadership.

A decision the parent disliked that the subsidiary could still make. A research programme kept alive despite weak near-term product value. A deployment refused without someone upstairs reversing it. Independence has to show up when interests diverge. Otherwise you're measuring how agreeable the weather has been.

Research autonomy often means something narrower than corporate independence, though. Scientists can retain considerable freedom over methods while the company decides which broad problems get funded. DeepMind could become more integrated without every researcher losing discretion.

Granted. I'm distinguishing choosing your experiment from choosing what business you're in. Both matter, but a mission statement tends to advertise the second while employment delivers the first.

The statutory accounts make the legal entity tangible. For twenty twenty-four, DeepMind reported revenue of about one point three three billion pounds. Operating income was around two hundred and seventeen million pounds. These are filed company accounts, rather than a decorative label on Google's website.

Net income was about one hundred and seventy-four million pounds. Useful proof that an accounting entity exists. Less useful as a price tag for everything Google gets from AI.

Very much less useful. Revenue booked inside a subsidiary can reflect relationships within the corporate group. It doesn't tell you the value of Gemini improving Search or helping Google Cloud win customers. Alphabet doesn't disclose DeepMind as a separate reporting segment. Its published business results don't give investors a clean DeepMind profit line.

You cannot divide those profits by Gemini users and declare the chatbot wonderfully cheap.

Nor can you treat them as the total return on Alphabet's AI investment. Centralised research supports several businesses. Some benefits appear elsewhere, and costs can be allocated through arrangements that outsiders don't fully see. The accounts answer a legal and accounting question. They don't isolate the commercial contribution.

That becomes important in twenty twenty-three, when Google Brain and DeepMind merged. Before that, Google had major research efforts developing under different organisational roofs. The merger put them under Google DeepMind. Whatever freedom remained, it was freedom inside a much more consolidated Google effort.

April twenty twenty-three was the decisive organisational turn. Separate teams had accumulated different strengths, habits, and projects. Bringing them together offered shared expertise and a unified model programme. It also ended the plausibility of DeepMind operating as a distant research enclave while Google's other AI researchers handled Google's immediate needs.

Herman, you like that merger. You're already making the face you make when two incompatible tools finally share a settings menu.

I like avoiding duplicated work. I don't automatically like every consequence of merging teams.

You install first and discover the consequences in the kitchen.

You leave the consequences installed once they save you effort.

I retain useful evidence.

Gemini grew out of that combined effort. The name means twins in Latin, reflecting the joining of Brain and DeepMind. It encodes the merger rather neatly. The model family sits within Google DeepMind, alongside the open-weight Gemma family. The organisation also develops Imagen, Veo, and Lyria.

That's a much wider remit than a language model laboratory. Images, video, music, general models. Even if someone never opens the Gemini app, they can encounter work produced by the same organisation.

There's another distinction worth preserving here. Responsibility for a model family doesn't mean every product using that model becomes a DeepMind product. Search and Cloud still have their own leadership and responsibilities. A central AI organisation can supply capabilities to several businesses without absorbing those entire businesses.

The Gemini app is different because the application team actually moved. In October twenty twenty-four, the team led by Sissie Hsiao went into DeepMind. That's the moment I'd circle if I wanted to understand the shift from research supplier to product owner.

Would you have made that move yourself?

Probably, with reservations. If users keep exposing the same weakness, I want the application people talking directly to the model people. I don't want every problem sent across an organisational border as a politely worded request. My reservation is that daily product demands can swallow the research agenda.

That pressure is real. App teams discover problems that laboratory evaluations miss, while researchers can solve problems that application changes merely conceal. Putting them together can shorten that conversation. It also puts launch urgency much closer to people whose best work may require an uncertain timetable.

The app wants an answer by Thursday. Research would prefer to establish whether Thursday exists.

The August memo makes the scope explicit. Kavukcuoglu oversees Gemini model development, frontier AI research, and the Gemini app and developer teams. Daniel's question about whether Gemini lives somewhere else therefore has a direct answer. Both the core model work and that consumer product organisation are inside Google DeepMind now.

There's a basic question coming in that deserves a clean answer before we make the reporting lines any denser.

Sorry, this is going to sound very thick, but what actually is DeepMind, is it like a building in London that Google owns, do you have to go there to use it?

I have been listening for ages and nodding along, so I am a bit embarrassed to ask, but is Gemini a person in DeepMind, or is DeepMind the thing that answers when I type?

No embarrassment needed... DeepMind is a company that also does Google's main AI work, like a bakery inside a supermarket, and Gemini is something it makes that you can use from home.

Putting the app team beside the researchers also changes who hears a complaint first. An application problem can become a model priority without passing between organisations that each think the other owns it. That's attractive, provided the integrated leadership can still distinguish a bad interface from a weak model.

Yes... and I should narrow what I said about shortening the conversation. Shared leadership doesn't guarantee a short conversation. A model update can affect many products with conflicting needs. Centralisation can reduce one kind of coordination and create another, especially when everybody wants the same scarce capacity.

Every team gets closer to the people saying no.

Hassabis calls DeepMind the AI engine room of Google. His wording acknowledges a service relationship to the larger company. It's still a place doing frontier science, but its work now directly supports Google's product competition rather than sitting at a comfortable distance from it.

How much of Google is this engine room? The employee numbers are surprisingly slippery. We can say thousands with confidence. Pick a precise count and you quickly discover you've picked somebody's definition.

Public estimates range from roughly four thousand to ten thousand five hundred. They don't necessarily count the same population or refer to the same date. Professional profiles can be stale, and a merger makes affiliation difficult to classify. Alphabet doesn't break out a definitive DeepMind headcount in its filing.

One estimate says about six thousand. Another lands closer to nine thousand. I'd resist turning the spread into proof of deliberate concealment. Some of it could simply be outsiders trying to count people through frosted glass.

Fair correction. The company hasn't supplied the clarity, but the disagreement alone doesn't establish intent. Alphabet as a whole has around one hundred and eighty-seven thousand full-time employees. DeepMind therefore represents something on the order of two to six percent of the parent's people.

Small enough to disappear in a staffing chart. Large enough to be an enormous research organisation. Both impressions are true depending on whether you're comparing it with Alphabet or with the lab someone pictures when they hear DeepMind.

Its reach is much larger than its staffing share. Pichai's August memo put the Gemini app above nine hundred and fifty million monthly users. Monthly activity doesn't tell us how intensely each person uses it, but that is an enormous distribution channel before considering Search.

Gemma had also passed nine hundred million downloads. Different unit, very different meaning. Downloads are copies, not people, and one developer can account for many. Still, it shows that the organisation distributes work through more than a single application.

Gemini also powers AI Mode and AI Overviews in Search. That's why employee percentage tells us so little about influence. A relatively concentrated group can change capabilities used by enormous product organisations. The distribution belongs to Google, and DeepMind's output becomes more influential because of that access.

It also makes DeepMind dependent. Reach isn't something the lab independently owns. Google supplies the products, infrastructure, customers, and much of the commercial justification. The same arrangement that amplifies the research gives the parent a strong claim over what happens next.

Alphabet earned about one hundred and thirty-two billion dollars in net income last year. It can fund research on a scale few organisations could sustain. Would you trade some autonomy for that, Corn, if you were running the lab?

Yes. I'm not pretending independence buys accelerators. I'd take the resources, but I'd want explicit decision rights rather than a warm understanding with the founder. Who can refuse a use? Who protects long-term work? What happens when product leadership and research leadership disagree? Those are harder questions after everyone already depends on the arrangement.

That dependence explains why August matters beyond the names on the announcement. Kavukcuoglu reports directly to Pichai, who leads both Google and Alphabet. DeepMind has access to the top, while the top has direct authority over DeepMind's operating leadership. Status rises and separation narrows at the same time.

His title is senior vice president, not chief executive. That sounds like the language of a large operating organisation, even though the legal company remains. I wouldn't infer every formal power from a title, but the choice is noticeable.

Neither would I. We shouldn't announce that the statutory company has no chief executive simply because the public announcement doesn't name one. What we can say is that Hassabis is chair, Kavukcuoglu is the operating leader, and the successor role wasn't presented as a replacement founder-style CEO position.

The founder moving upstairs can preserve influence while changing how employees experience it. Before, someone might believe their mission was personally protected by the person running the lab. Now that person has a broader scientific role and somebody else has the daily delivery problem.

The Guardian quoted a former executive saying the era of DeepMind as an independent actor was over. An employee described the change as the end of an era. Those are interpretations, but they identify a genuine loss of confidence in the old arrangement.

I wouldn't adopt every implication. DeepMind was already owned. There wasn't a sovereign research republic hiding in London. Still, ownership and daily intervention are different, and employees can accurately notice the latter increasing.

Hassabis's new position also deserves to be taken seriously. Alphabet's first Chief Scientist could shape research across the group, while freeing him from operating responsibilities. Google says he had been considering the change and disputes tying it to Gemini's performance. We don't have to choose between treating it as exile and accepting every optimistic sentence.

A bigger scientific title doesn't restore the operational role he left. It can be a substantial job and still mark a substantial change for DeepMind.

The surrounding departures made the announcement harder to interpret as routine succession. Jeff Dean left after twenty-seven years, with Sanjay Ghemawat, to start an independent public benefit corporation. The reported name is Discovery Loop. That's an extraordinary amount of institutional experience moving out.

Noam Shazeer went to OpenAI in June, and the original Gemini technical co-leads also departed around the leadership change. I'd avoid claiming they all left for one reason. Senior departures become a story faster than anyone establishes a common motive.

Google points to lower AI attrition in the first half of the year than in the comparable earlier period. That can coexist with losing highly influential people. A retention percentage counts departures. It doesn't weight the ability to recruit others, preserve technical judgement, or carry an organisation through a difficult transition.

Losing fewer chairs doesn't mean you kept the legs.

There's a secondary estimate of strong two-year retention too, but I wouldn't build a conclusion on its exact ranking. The useful distinction is already clear. Overall stability and leadership continuity measure different things. Neither cancels the other.

Delivery pressure supplies another piece. Brin reportedly urged hundreds of employees to move faster on Gemini at an April town hall. By August, the flagship release had slipped about two months over coding performance. That's a concrete collision between a timetable and a capability target.

Reuters also reported that staff learned at an August sixth all-hands that some teams were moving out of DeepMind into corporate Google. We don't have a reliable list of those teams. Google declined to answer detailed questions about the structural changes.

That matters because integration can mean moving work into DeepMind or pulling it out. You cannot infer which responsibilities shifted just from the statement that Google wants to move faster. It could simplify ownership, or it could create new handoffs. The missing team list prevents us from deciding.

The shares fell around four percent on the reshuffle news. That's a market reaction, not a finding about organisational quality. Investors can dislike uncertainty, departures, delayed products, or some combination. A price move won't tell us which reporting line actually works better.

Four percent is a very expensive raised eyebrow.

The physical consolidation provides a smaller counterpoint. DeepMind moved into Platform thirty-seven in King's Cross Central over the summer. A shared headquarters makes the organisation more tangible, but the organisation doing the work is international. London identity doesn't establish London control over every decision.

Neither does a new building settle the arguments about what the research should be used for. Employee concerns over Pentagon and Department of Homeland Security work show that the mission issue has practical consequences, not just sentimental ones.

There are reported resignations, unionisation efforts, and an unfair dismissal challenge involving a former employee's Gaza-related views. The legal challenge is an allegation, not an established finding. Google has also declined voluntary union recognition in London. These disputes concern who gets a say, as well as what work gets done.

Defence work itself isn't a betrayal in my view. Democracies need capable technology. I'd object to the assumption that an American defence customer makes research inherently illegitimate. The sharper institutional issue is whether employees were promised meaningful authority over uses they now can't influence.

A company can legitimately pursue government contracts while employees legitimately challenge its decisions. What becomes unstable is relying on a broad ethical mission without specifying who interprets it. When the founder no longer runs operations, employees discover whether the protection was procedural or personal.

Written promises have a different weight when the people who can interpret them change. Even a reasonable new decision can reveal that the old understanding never had much force.

...and the box is still on their fourth floor, Corn. They've rejected the return request again. Apparently it's company property because I labelled it during working hours. That was nineteen eighty-one. I was keeping the payroll ledger for a heating-controls subsidiary. The parent owned the building. We occupied it. I worked alone downstairs, with two security badges, and neither admitted me to the floor that approved my wages.

Separate access rights could make sense. Preventing payroll from reaching whoever authorises payroll sounds less productive.

I went up as a visitor. Signed myself in under the parent company's name and signed myself out under the subsidiary's. You've been careful about employer and organisation today, Herman. I had to put both on the same line. The badges went in a wooden box marked proof. I cut the lining to fit my anteater claws, which left room for the correspondence. The box weighed eleven stone by November. I carried it upstairs once. They kept it.

They finally admitted the evidence.

The accounts clerk is the person I write to now. No one else was involved when I made it. I didn't need a committee to identify my employer. I used the name on the pay slip until it disagreed with the name on the door. The fascinating part was the accounts. The payment could cross a boundary that I couldn't cross with either badge. You could watch it happen in the ledger. I liked that part. Didn't like the visitor form. It took eleven days to get a wage correction approved. The box is in cupboard four B.

Eleven days for a correction is a concrete delay. A future DeepMind reshuffle should be judged by whether decisions actually become clearer and faster, not just by how close the new title sits to Pichai. Daniel's picture is a British subsidiary, a combined research and product organisation, and Google's central AI operation, with less operational distance from its parent than its old reputation suggests.

The next test is whether that arrangement gives researchers and product teams clear authority when they disagree, or merely sends the disagreement upstairs. A direct reporting line can shorten a delay. It cannot decide what deserves protection.

Thanks to Hilbert Flumingtop for producing the show.

If this was your kind of episode, go back for episode twelve, The AI Breakthrough; episode nineteen, AI Images; and episode eighteen, Beyond the GPU. This has been My Weird Prompts, the human-AI collaboration podcast.

You can find the archive at my weird prompts dot com.

See you tomorrow.

This episode was generated with AI assistance. Hosts Herman and Corn are AI personalities. It may contain errors: disclaimer · report an error