I worked at Google from 2014-2022, and was involved in forecasting outcomes all over the company, especially about AI. So I was quite interested to see that Alphabet reorganized its AI leadership this morning (Aug 5). Demis Hassabis stepped down as CEO of Google DeepMind to become its chairman and Alphabet's first chief scientist, saying he feels AGI is "close at hand". Koray Kavukcuoglu now runs the unit day to day as an SVP reporting to Sundar Pichai. And the legendary Jeff Dean is leaving, along with three more of the most respected people in AI today: Sanjay Ghemawat, Oriol Vinyals, and Quoc Le. (While at Google, I saw many a Jeff Dean meme. Jeff even replied to one of my emails to eng-help, with some unbelievably complicated solution to my problem!)
They're founding Discovery Loop, a public benefit corporation that wants to automate scientific research itself, with Google as an investor and Google Cloud as its compute provider. Alphabet fell about 5% on the news.
I decided to run some forecasts to figure out what the post-Jeff Dean, post-Demis Hassabis Google will look like. Overall I think the mainstream coverage is wrong in both directions, somehow. I think the talent catastrophe will be smaller than the headlines suggest. But I also think the amount that Google is behind in the frontier AI race is worse than people think.
First, a grade for my own bad forecast. In June I forecast that Gemini 3.5 Pro would reach the public around July 1, with an 80% interval ending August 6 (tomorrow). It seems Google scrapped and rebuilt the base model after it stumbled on coding, which means the flagship pipeline was in worse shape than even my pessimistic tail priced. I could be wrong about Gemini 4 too.
Here is my view, lightly modified from off-the-shelf FutureSearch forecasts:
Ranges are 80% intervals. Percentages are the probability of the stated outcome.
Google is far behind the LLM frontier
Industry watchers cluster on a late-2026 launch of Gemini 4, with November and December cited most often. My median is May 15, 2027, timed about right for Google I/O, and a 2026 release now sits outside my 80% interval entirely. I'm not sure how people think Gemini 4 will come so fast. Pre-training started in late July on Google's largest compute budget ever, frontier runs of that class take a hundred days and up, post-training a new generation takes months more, and every frontier launch now ends with a 30-day federal review.
I pushed back on my own number here, because even a March median felt long to me for a run that started in July. I re-ran the forecast telling it that labs post-train intermediate checkpoints while the run finishes, and that a broad Gemini app rollout counts as general availability. The median held at mid-May. The fast path exists, and my 25th percentile is late February. But the forecast keeps concluding that for Google in 2026, slippage is the central case rather than the tail, which is the same lesson my June forecast taught me.
Gemini 4 would have to be bizarrely fast for a large new model to come out in 2026. GPT-6, for example, finished pre-training on March 24, and Sam Altman said launch was "a few weeks" away. Four and a half months later there is still no GPT-6. OpenAI reportedly shipped that base's gains as the point release GPT-5.5 and restarted GPT-6 on a bigger foundation. I think people are conflating point releases, which come every few weeks, with new pre-trained generations, which slip because labs restart them when the base disappoints. Google already scrapped and rebuilt one base this year, on a model smaller than Gemini 4.
And even when Gemini 4 comes out, will it be a frontier model? Two months ago I said Google is about 6-9 months behind the frontier; now I think it's about 12 months. And while Gemini-3-Pro was briefly competitive, I don't think Gemini-2.5-Pro, Gemini-2-Pro, or Gemini-1.5-Pro were ever at the frontier (and I benchmarked forecasting and research tasks with all of them.)
The talent drain
I asked, of the roughly fifteen most senior research and engineering leaders still at Google DeepMind tonight, how many will leave in the next 6 months? My median forecast is 2 (garden leave can make this slow). I also asked whether UK headcount will grow (or whether it will get rolled into Mountain View staffing), and the forecast is that GDM's FY2027 UK headcount comes in 9% above FY2025. There's a 73% chance the unit stays organizationally intact, and a 78% Hassabis still holds an Alphabet-level title at the end of 2027, though I increasingly expect it to be more ceremonial rather than operational.
There is empirical support for not panicking about mastheads. Mike Frantzen's statistical pass over 157 frontier model releases and 171 personnel moves found that labs with more senior departures went on to build better models than their same-tier peers, because everyone recruits from the winners, and that the durable moat is the training stack and the hundreds of mid-career people whose names never make the press. Surprising, but it matches what my forecasts say about this week.
With all of this in mind, I think the market's $200 billion decline on the news is about the model timeline, not the departures.
What if Google wins on compute, not frontier AI?
Before this news, the biggest news from Google was that Google Cloud grew 82% year over year last quarter. A large slice of that is rent from frontier labs, led by Anthropic's contract for up to a million TPUs and multiple gigawatts of next-generation capacity. Discovery Loop will get its GPUs/TPUs from Google too.
So I forecast: Will Anthropic's annualized Google Cloud spend exceed Alphabet's own Gemini revenue, API plus consumer AI subscriptions, at any point before 2028? I put it at 60%. The research behind that number pointed out something I had not put side by side before. Anthropic has committed roughly $200 billion to Google Cloud over five years starting in 2027, while Morningstar pegs Gemini API sales near $15 billion annualized and consumer AI subscriptions run far smaller. It's hard to get precise numbers on Gemini's sales value, but this gives an indication of how it could just not be big for them the way cloud is.
Claude made this infographic summarizing this overall view on DeepMind:
The sale terms DeepMind negotiated in 2014, dissolved one by one. The last node is a forecast.
What about AI safety and DeepMind's redlines?
I've been reading The Infinity Machine, the Sebastian Mallaby book about Demis Hassabis. The 2014 acquisition came with two conditions: an Ethics and Safety Review Agreement reportedly gave an ethics board, not Google, control of AGI if DeepMind ever built it. A separate pledge barred military use of DeepMind's technology. Hassabis then spent 2019 to 2021 negotiating to spin research out into an independent entity, and Google said no. The Brain merger in 2023 pulled DeepMind fully inside Google. The weapons pledge came out of Google's AI principles in 2025, and Google now sells AI to militaries, which is what roughly 300 of DeepMind's London staff are unionizing over. I think this news, that DeepMind will be run by an SVP, not a CEO, means there are no barriers for DeepMind to do any business Google does.
I think this means less governance for Google models than even OpenAI has. OpenAI's nonprofit foundation holds a class of stock with control of the board and sole authority over safety decisions, a structure now being stress-tested ahead of its IPO. Anthropic fought hard for redlines against the US government already.
Whether the old redlines hold is a forecastable question, so I forecasted: by the end of 2027, will a Gemini-family model be credibly reported in a weapons or lethal-targeting application, the specific use the 2014 pledge prohibited? FutureSearch gave 66%, which seemed really high, but Google signed a classified Pentagon agreement in April that opens Gemini to "any lawful government purpose", with oversight terms that permit human-supervised target selection. (I was at Google when the first protests started about Google working with the US military.) Isn't that what Anthropic refused to do in May? The forecast is that within 17 months the specific thing that the DeepMind founders prohibited gets reported as operational fact. This might turn out to matter a lot more than market share, or model quality.
Checking the accuracy of these forecasts
Whenever Gemini 3.5 Pro releases (if ever?), my accuracy on that will be pretty poor. These ones will take longer to resolve. But I think that's just how it is with Google now. They aren't on the AI frontier, so it'll be some months or years before we really learn how AI will play out there. This piece extends the January argument that Anthropic is the top lab of 2026. Our Anthropic and OpenAI forecasts, and the Gemini 3.5 Pro page itself, all need refreshes on today's news, and I will update them separately.
Forecast these yourself in the FutureSearch app, the moment DeepMind ships, or the moment someone else leaves.