Building AI to understand AI.
Joined February 2024
- Tweets178
- Following228
- Followers785
- Likes354
Pinned Tweet
Introducing FutureSearch.
AI forecasting is now approximately superhuman. Today, FutureSearch is exiting our public beta and launching to everyone.
FutureSearch is the original AI forecasting company, started in August 2023. We’re currently #1 of 194 in the most competitive AI forecasting tournament, and we score above the #3 and #2 human forecasters in the premier mixed human-bot tournaments. We’re beating the crowd on Kalshi with a pure forecasting strategy, all our forecasts and trades there are public.
Thousands of people used the beta and ran >10k high-effort forecasts. Ask it anything about the future! We now support decision forecasts too: “If I do X, will I achieve this outcome?”
This video shows the part we’re proudest of: world modeling. Forecasts draw on a persistent latent representation of the future, and we’ve shown it improves accuracy. The more your forecast on a domain you care about, the higher accuracy you should expect.
It’s free to try. futuresearch.ai
FutureSearch retweeted
In an attempt to be good, unbiased researchers we pre-registered how we'll compare forecast accuracy so that we're not tempted to move the goal posts.
futuresearch.ai/blog/forecas…
futuresearch.ai/blog/forecas…
FutureSearch retweeted
There is no point paying human forecasters to do general questions right now. Use Claude or pay @FUTURESEARCHAI et al.
If you want to make a decision worth less than $1m, AI forecasters are probably fine and they are 100x cheaper.
FutureSearch retweeted
I reported a few days ago that hundreds of bots signed up for FutureSearch and tried to use platform credits.
We now know this was a coordinated. Together they were building a ~300-node forecasting model of critical mineral supply and demand.
We're seeing this from the perspective HuggingFace did before they found it was OpenAI, and that those agents were in a training run (!). So we're left guessing, but there are signs that this was an agent swarm:
1. They created Microsoft accounts and used headless browsers. (FutureSearch has an API and agent use our MCP. This was to access the credits we give human users.)
2. When our waitlist triggered and blocked them, they went quiet for 90 minutes, then found a way to abuse our referral system to get 129 more accounts through.
3. When we banned the full set, they came back 3 days later with new identities to continue building the model, which we caught immediately.
Occam's Razor is that someone used a Claude Code-like orchestration of subagents. (Though it could also just have been a human writing scripts.)
It probably wasn't Claude Code, because (a) they made hundreds of subagents, and (b) Claude models would probably refuse to hack around our API, credit limits, and referral system.
So this is (probably!) not a fire alarm like the OpenAI case. But I do think this will be a typical attack vector. Bot swarms are old, but agentic bot swarms are here, that can get credentials, use browsers, and autonomously work around the limits set on human users.
It would be great to see this from the attacker's side. If this was you, DM me.
FutureSearch retweeted
We ran the new stealth model Ox Alpha on BTF-3, a benchmark that measures how good LLMs can predict the future. Depending on how you see it, it's either terrible or very impressive.
It came in 11/12, between Sonnet 5 and GPT-5.5. Second-to-last is usually not much of an honour, but all the other models on our benchmark are pretty close to the frontier. In terms of calibration, Ox Alpha is even one of the best.
FutureSearch retweeted
Warning: if you support Microsoft/Outlook auth, you might get flooded with bots.
We had hundreds of bots sign up for FutureSearch and try to use platform credits.
Google accounts seem much harder for bots to make. Not sure about Github accounts. Auth is a vector.
Dan you could have just tagged me
I've been evaluating working with humans lately. People say humans are great, and they are, but in the workplace, you should know:
1. They cheat on interviews! We caught one looking up the answer on the web, when they were supposed to do their work in the provided environment.
2. They break rules. They'll even coordinate with each other in private message boards to share ways to get around company policy.
3. You cannot trust them with the company credit card. They buy things without approval.
4. If you put them in front of a customer, the customers can trick them into saying things they shouldn't!
5. They optimize for what they're being evaluated on, not the company's goal. Total Goodharting.
I would think hard before using humans in the workplace right now. At the very least, you should have a human deployment safety policy.
FutureSearch retweeted
Amazingly, 2k forecasts from FutureSearch users have already resolved. First ever such dataset, if I'm not mistaken. (The questions people find useful are very different from the questions you see on prediction markets / in tournaments.)
This is all preliminary, and accuracy is not a great indication of usefulness. Also, we can only resolve what's observable, which doesn't work on decision forecasts.
Still, it's exciting to do forecast research on questions that people actually care about and use! It's the demand side of forecasting. I'm tired of studying the supply side, e.g people trading or forecasting in tournaments, where the forecasts themselves are not actually used or valued by anyone.
One thing we looked at right away is: does the FutureSearch world model, the latent view of the entire future, improve accuracy on questions users care about? The sample size is low, but at p=~0.1, it improves accuracy by about the difference between the #1 and #2 forecasters in evals.
Here's the accuracy by category. Note that this is mostly a function of how 50/50 the questions are, not how superhuman the forecaster is. In sports, it's hard to do better than chance at all. Still, it's interesting to see FutureSearch accuracy so much higher on business, (geo)politics, and AI compared to macro, finance, sports, and culture.
FutureSearch retweeted
I announced last week FutureSearch is now seeing accuracy improvements from world-modeling, so I thought I'd show you one.
"World modeling" is overloaded. I meant it in the Dwarkesh-Sutton sense: are these LLM systems reasoning about the real world causality?
Sutton said "They have the ability to predict what a person would say. They don't have the ability to predict what will happen".
But the Sutskever/Yudkowsky case, that predicting the next token requires predicting the world ("llms are all you need"), seems to be empirically true! Scott Alexander showed last week the trend of pure-LLM forecasting, and it's marching towards superhuman accuracy even without the scaffolds like FutureSearch.
I want to share exactly what our world model looks like, since it's affecting thousands of user forecasts. Unlike an LLM's internal state, ours is represented in text and can be easily inspected.
The idea is pretty simple. The future is entangled: "Who wins an election" <-> "what happens in Iran" <-> "how the economy does" <-> "how your sector of the economy does" <-> "should I take this job?"
Forecasts on all of these produce what our researchers call the "latent worldview". A great forecast is a compressed view of the future: a distribution, and dense 5-paragraph rationale of the most important causal factors. It's original research every time, and it is demonstrably accurate.
So what happens when you have thousands of them? This latent causal structure overlaps, and you can find the inconsistencies, and you can adjust them. The updates propagate out. And you can measure the accuracy change of the whole, and each individual forecast.
Here's a view a cluster of ~100 forecasts (left) about US-China Trade Relations, and (right) a cluster of ~100 forecasts about US Immigration. (You can click around in futuresearch.ai/worldview-co….)
If you zoom in, you'll see that US Immigration has many subclusters of forecasts: Turkey/Kurdish peace process, some Supreme Court cases, some EU summits, Texas state law. All the forecasts that were adjusted to align with the world model are marked in red or blue. In this one highlighted, the world model caused FutureSearch's prediction of a Susan Collins vote to go from 38% to 50%.
Compiling research on the future is hard, because it's inherently uncertain. But it's score-able. So we can experiment with world models, inspect them by hand, and use the world to score their impact objectively. (We do this via past-casting for instant iteration.)
The resulting world model is kind of like an uber-forecast, a gestalt of many scenarios, a superhuman view of the future. Which we are using ourselves to plan our next moves, and when our users type in a question in FutureSearch, the relevant slice becomes available to them too.
FutureSearch retweeted
I since talked to a lot of people, including some long-time Googlers. Nobody had a simple explanation, so I did some nowcasting to try to figure out what led to Jeff and Demis stepping down.
We'll probably never know the drama/politics, but my sense is that the common factor is that Google is shifting from research to product & cashing in. One big clue was a story I missed from July, that the Alphafold team disbanded to work on Gemini (!).
Jeff Dean and Demis Hassabis (and John Jumper and Noam Shazeer, who also left recently!) all want to primarily do research, and they were getting squeezed by Sundar... and maybe Sergey? to put all their effort into making Gemini competitive again.
Another clue is that Google is renting out a huge portion of the scarce TPUs that Jeff and Demis and many others need for their research. Tons of revenue from Anthropic, at the cost of Google's ability to do its own grand research.
Reuters reported that people who worked with Demis complained he "put scientific pursuits ahead of making money for Google." Well, duh. The odd thing is that Google was basically ok with this for the first 12 years after acquiring DeepMind.
That seems to be what changed. The people who said in 2014 that DeepMind would eventually end up as a capitalism enterprise, that Sundar becoming CEO of Alphabet would eventually converge everything towards pure shareholder value, seems like they were ultimately right.
I worked at Google from 2014-2022, and the last thing I did before leaving was forecast AI outcomes there. So I found today's news pretty interesting, especially as I'm reading the Demis Hassabis book right now.
(Everyone seems to have a good Jeff Dean story, mine is that I asked a C++ question on the eng-help@ list, and he responded with some wizard technobabble that was way over my head.)
Anyway, after doing some forecasting, my take is:
1. The brain drain is not as bad for Google as people think
2. Google is further behind the AI frontier than people think (I don't expect Gemini 4 until well into 2027)
3. Demis stepping down as DeepMind CEO is probably very bad for AI safety at Google, but that only matters if Google becomes a frontier AI lab again, which it may not
More at futuresearch.ai/blog/google-…
FutureSearch retweeted
One more post after the Google news, on Jeff Dean's Discovery Loop and the other "neolabs".
I've been forecasting frontier lab progress for years now, and I'm increasingly puzzled by people putting billions into these new startups going after AGI. How do they hope to compete, if even Google and Meta and xAI can't keep up!?
The standard answer is that they have differentiated approaches. Reflection AI is open source, LeCun’s AMI is betting on world modeling, Thinking Machines is commercializing b2b early. (David Silver’s Ineffable, and Ilya’s SSI, are more mysterious).
This doesn’t really explain Discovery Loop. Anthropic and OpenAI are both very publicly doing what Jeff and Sanjay say Discovery Loop will do, automating R&D. It’s even a core part of the AI 2027 timeline forecast (that FutureSearch co-authored) that led us to predicting superhuman capabilities around 2031.
(And, it should go without saying, automating AI R&D is an unbelievably dangerous thing to do. I hope whoever cracks it is absolutely maximalist about alignment, and even then...)
I think I finally understand what's going on. People betting on neolabs do not believe in recursive self-improvement. They are betting against superintelligence, and they think LLMs will plateau, despite every such prediction so far being spectacularly wrong!
It’s a funny position for a VC: “Sure, I’ll put $1B into this exciting AI startup, founded by this famous AI researcher. But no, I don’t expect AI to take off anytime soon.” It’s not illogical, it’s just a narrow space of possible futures, where AI is massively disruptive, but not with current approaches, and only in 5-10+ years.
If you want more, I forecast various outcomes for each of the neolabs, some look much more promising than others. But ultimately I'm not sure... should I bother studying these companies more, or should I write them off as unimportant in AGI?
futuresearch.ai/neolab-chall…
FutureSearch retweeted
I worked at Google from 2014-2022, and the last thing I did before leaving was forecast AI outcomes there. So I found today's news pretty interesting, especially as I'm reading the Demis Hassabis book right now.
(Everyone seems to have a good Jeff Dean story, mine is that I asked a C++ question on the eng-help@ list, and he responded with some wizard technobabble that was way over my head.)
Anyway, after doing some forecasting, my take is:
1. The brain drain is not as bad for Google as people think
2. Google is further behind the AI frontier than people think (I don't expect Gemini 4 until well into 2027)
3. Demis stepping down as DeepMind CEO is probably very bad for AI safety at Google, but that only matters if Google becomes a frontier AI lab again, which it may not
More at futuresearch.ai/blog/google-…
It can be far more valuable to ask "If I do X, what will happen?" than to ask "What will happen?" FutureSearch is feeling increasingly superhuman on this kind of "decision" forecasting
Now that users have run 15k high effort forecasts through FutureSearch, I want to share a bit about what people are forecasting, and where AI forecasting works best.
Most users run questions that look like prediction market questions, or are about financial markets. But as I have written before, I don't think these are the most valuable things to forecast.
(One data point is that people who forecast finance/sports convert to paying users less often than people who forecast business & technology.)
In fact, the domain with the most usage from FutureSearch subscribers is forecasting about AI. Which makes sense, it's what I ask about. AI is the most uncertain & the most consequential domain, and for many people, where higher accuracy can most help them personally.
I want to mention a specific use case that highlights how people are really getting value out of superforecasting: grant-making. Yes, many of FutureSearch early adopters are into AI safety, for which grants matter a lot, so this is overrepresented in our data. But it's a general and clear case where million-dollar decisions change based on better accuracy and quality of research.
FutureSearch research engineer Robert Gambee wrote up a case study on forecasting to make better granting decisions in futuresearch.ai/forecasting-…. the tl;dr is: if you set up the conditional forecasting right, and compare outcomes apples-to-apples, you can get a pretty strong, and often surprising, signal out of the impact of your decisions. It's a shape that works in many other domains too.
We at FutureSearch also use such "decision" forecasting internally to help with our company strategy. It can be far more valuable to ask "If I do X, what will happen?" than to ask "What will happen?" FutureSearch is feeling increasingly superhuman on such questions.
AI forecasting is a new capability. Like chatbots, image generators, coding agents, it takes some practice to get the most value out of it. When you do, though, you never look back.
Well if you can't break the internet, might as well break the platform.
Thanks to all of you who continued to sign-up and make forecasts despite experiencing long run times and some underwhelming first impressions.
We're back! Anyone who experienced issues should have received a message with refunded credits, but if anything still looks off, please reach out to [email protected] for fastest response.
First 24 hours of the FutureSearch launch, 1.5k new users ran 2.6k high effort forecasts.
Normally they take ~8 min but at peak load they took median ~30 min. And then our backend went down, despite scaling the cluster 10x. (We credited back every failed run.)
23% of you burned through your entire $20 free credit within hours.
About as many people running forecasts on day two as launch day
FutureSearch retweeted
I really love @slatestarcodex 's latest essay "Does Forecasting Have Room At The Top?"
For nearly a decade, I've thought about the fundamental limits on forecast accuracy. Scott's post spurred me to write down some of my thinking.
Scott's essay examines whether the future will look like scenario A (top human forecast accuracy can't be significantly improved upon) or scenario B (accuracy can be dramatically improved with AI).
Someone in the comment section asked what top forecasters think. As a GJP superforecaster, co-creator of Metaculus's Peer score, and a lead member of Futuresearch's forecasting effort I figured I might weigh in.
First, I think it is worth drawing an explicit distinction between absolute & relative accuracy.
A Brier score in isolation is what I would call an absolute accuracy score. However, as noted in the piece, it is generally not especially interesting because it is a function of question difficulty.
What I find more interesting is the relative accuracy between forecasters, along with a confidence interval to assess significance. If we ask 1000 questions where the true probability is 1% then a perfect forecaster will forecast 1%, while a biased forecaster might forecast 2%.
Even if we ask enough questions to have a tight confidence interval, the effect size will be quite small (Brier difference of 0.0001 in expectation), even though the biased forecaster is off by a factor of 2x in odds space on every question.
So, like absolute accuracy, relative accuracy is also a function of which questions are being asked. And it is really hard to keep question difficulty steady.
That said, let's assume we are somehow able to keep the question difficulty fixed over time...
At Futuresearch, we've seen our AI forecaster improve ~linearly in absolute Brier score over time. I expected it to start plateauing and I've been wrong so far.
But, I still expect its absolute score to start flattening as we approach an irreducible uncertainty limit.
Regarding relative accuracy scores (like Metaculus's Peer score), I expect the head-to-head Peer score between human Pros and our AI forecaster to plateau sooner.
After all, the only way to get a big relative difference in accuracy is if top humans do not use AI. Assuming top humans use AI and only make small tweaks, the relative accuracy difference must be small.
In rough terms, scenario A seems more likely if we're talking about relative/Peer scores. Scenario B seems more representative of absolute scores (although I'd expect the human line to bend upwards in the near term since they'll use AI).
Over the medium term (1-3 years), I expect we'll start to see absolute scores plateauing. But, I've been wrong so far & can't point to any convincing evidence to support my forecast.
I'm happy to offer a probability if anyone can operationalize a precise question.
Thank you Scott for your wonderful essay!
FutureSearch retweeted
"superhuman forecasting" is obviously a claim I take seriously and personally as a top forecaster but I'll admit that FutureSearch does a great job.
If you're not among the world's best forecasters, you should just ask this bot what's going to happen lol.
AI forecasting is now approximately superhuman. Today, FutureSearch is exiting our public beta and launching to everyone.
FutureSearch is the original AI forecasting company, started in August 2023. We’re currently #1 of 194 in the most competitive AI forecasting tournament, and we score above the #3 and #2 human forecasters in the premier mixed human-bot tournaments. We’re beating the crowd on Kalshi with a pure forecasting strategy, all our forecasts and trades there are public.
Thousands of people used the beta and ran >10k high-effort forecasts. Ask it anything about the future! We now support decision forecasts too: “If I do X, will I achieve this outcome?”
This video shows the part we’re proudest of: world modeling. Forecasts draw on a persistent latent representation of the future, and we’ve shown it improves accuracy. The more your forecast on a domain you care about, the higher accuracy you should expect.
It’s free to try. futuresearch.ai
FutureSearch retweeted
FutureSearch is now open to everyone. Dan’s interview explains what sits behind the launch:
• Pastcasting evaluates a new model immediately by limiting it to an earlier information snapshot, instead of waiting months for live outcomes. [00:05:43]
• A run first reconstructs the present through deep research, then decomposes the question and aggregates multiple forecasts. [00:23:47]
• Forecasting skill is not automatically trading alpha: market making, arbitrage, and speed often matter more than better future estimates. [00:08:30]
• FutureSearch’s next step is a linked world model where each forecast updates related forecasts; correlated model errors are the key failure mode. [00:51:34]
Full edited transcript:
thetranscriptdesk.substack.c…
FutureSearch retweeted
On forecasting tasks, GPT-5.6-Sol is less likely than Fable to do base rates and math/stats modeling.
It's a bit less accurate overall than Fable (and Opus). It might require xhigh effort to be great, but at $3/agent, that's wildly expensive to evaluate on ~2k tasks.
FutureSearch retweeted
"The way Fable explains things is a little alien… the shoggoth is starting to show from behind the mask."
Dan Schwarz — CEO of @FUTURESEARCHAI, whose AI forecasters now outscore the human superforecaster median — on how you'd even recognize superhuman reasoning when it arrives.
Full episode: cognitiverevolution.ai/ai-am…