Please add questions for what will happen in 2026 related to AI! I've added some clarifications below. If there is ambiguity I will resolve to my best judgement.
Clarifications
"SSI will release a product": It should be generally available in 2026; i.e. no waitlist. Should be an AI product; I’m not counting hats, clothing, etc.
"X will outperform the S&P": As measured at the end of the year. It's not sufficient for X to outperform at some point in the year.
"An LLM will beat me at chess": See this market:
“Epoch AI will estimate that there has been a training run using more than 5e27 FLOP” : according to this source or some other official announcement by the org.
"The METR time horizon will exceed X hours": At 50% success rate, acoording to this source.
“Frontier Math Tier X >= Y%” refers to the top score on this leaderboard. The current top scores as of 2025-12-21 is 40.7% for Tier 1-3 and 18.8% for Tier 4.
“An open millennium prize problem is solved, involving some AI assistance”: refers to these famously difficult mathematics problems.
“Epoch Capabilities Index >= X” refers to this metric. The current leader as of 2025-12-18 is 154.
"Yudkowsky will publish a new book": It should be avalible to (pre)order some time in 2026. Resolves no if they announce it but you can't (pre)order it somewhere.
"Metaculus will predict AGI before 2030" means that at some point in 2026, this market has an estimate before 2030. The current estimate as of 2025-12-24 is May 2033.
"There will be an AI capabilities pause lasting at least a month involving frontier companies": This means that there is an agreement between frontier companies not to advance capabilities. Frontier companies not releasing new models for a month does not resolve this market.
"My median ASI timelines will shorten": My current estimate as of 2025-12-27 is March 2033.
"My P(doom) at EOY (resolves to %)" : Resolves to a %, rather at the end of year, except in the unlikely case that it's 0 or 100%, in which case I'll resolve NO or YES. My current estimate as of 2025-12-27 is 25%.
"An open source model will top Chatbot Arena in the 'text' category": This refers to this source.
"US unemployment rate reaches 10% due in part to AI": Based on this source, according to my judgment.
"Schmidhuber will continue to complain about people not citing his work properly", e.g. here.
“Thinking Machines will train and release their own model”: from scratch, not a finetune of another model
AI Summaries below:
Update 2025-12-19 (PST) (AI summary of creator comment): "Open source model" is defined as a model where the weights are publicly available.
Update 2025-12-20 (PST) (AI summary of creator comment): "I will think that a Chinese model is the best coding model for a period of at least a week": Cost and speed will not be considered unless they make the model difficult to use. Resolution will be based on how well the model performs on difficult coding tasks encountered by the creator.
Update 2025-12-21 (PST) (AI summary of creator comment): "An LLM will beat me at chess": The creator is rated approximately 1900 FIDE.
Update 2025-12-25 (PST) (AI summary of creator comment): "A significant advance in continual learning": The model should be able to remember facts or skills learned over a long period of time like a human. It should not make egregious errors related to memory that current bots in the AI village regularly commit.
Update 2025-12-30 (PST) (AI summary of creator comment): "There will be an international treaty/agreement centered on AI": Must be signed by 3+ countries and be legally binding. Non-binding declarations (like the Bletchley Declaration or Seoul Declaration) do not count. Examples that would count: Montreal Protocol or Non-Proliferation Treaty.
1,000People are also trading
@mr_mino alright nw. lmk if you want to bet more, i'd def expect no for the simple reason that historically they have released new versions slower than once per year, and claude 5 isn't out yet
Related to the question: I thought everyone should see this: The members of the AI futures project have given an update and they appear to now be relying on the 80% time horizon length graph from METR for their predictions rather than the 50% time horizon length graph and their timelines have gotten longer. Here is there most recent update: https://open.substack.com/pub/aifutures1/p/ai-futures-model-dec-2025-update?r=6lp84s&utm_medium=ios. Daniel and Eli think that the scenario as laid out will basically be the same except that it will happen a few years later. Daniel pointed out that things appear to be 1 year behind the original scenario. As of right now Daniel thinks things will start to unfold in 2029 while Eli thinks things will start to unfold in 2030.
@mr_mino Unless I'm missing something on my own ChatGPT account, it hasn't been available to free users since the release of GPT-5. I would've kept using it otherwise because GPT-5.x-Instant is some hot garbage.
@mr_mino Is this the mobile app? Do you have legacy models for the free tier on it? (I'm only using the web version, even on mobile.) Weird that it's available there but not on web.
@moozooh my bad, this was the ChatGPT Plus subscription. Didn’t realize I had that. I’ll check again once I revert to free tier.
@mr_mino Wouldn't it be more correct to resolve N/A since the question wasn't valid in the first place? If it wasn't free at the start of the year, it can't "remain" free at the end of the year.
@moozooh You’re right that I was confused about my own account status, but GPT-4o actually was free from May 2024 until GPT-5 launched in August 2025. So the premise was valid even if my personal reason for thinking so was wrong. NO still seems like the right resolution.
@notadiron Good question. It should be signed by 3+ countries and be legally binding. For example, things like the recent Bletchley Declaration and Seoul Declaration don’t count, but something like the Montreal Protocol or the Non-Proliferation Treaty would count.
(sorry duplicate)
@Dulaman This is interesting but I'm not sure what relation it has to the tagged question. Decentralized training is still more than 2 OOMs behind the frontier and isn't on track to catch up by EOY 2026 assuming thier trendline.
@mr_mino fair point. I can think of three scenarios:
if we're in a really agressive scenario with China going all in on distributed training runs, then we may start seeing these go above 1e27 FLOP in 2026.
For Chinese companies all doing open source models. In theory it's in their interest to pool their resources together in a massive training run and collectively benefit from the better resulting model, instead of trying to compete with centralised training runs. But for that change to occur in 2026, definitely seems like a stretch.
@Dulaman I agree 5e27 FLOP seems unlikely in 2026, but I can add a question like “Largest distributed training run exceeds 1e27 FLOP” if you’re interested in betting on a lower value.
some sources suggesting that Chinese companies have been derisking the distributed training option:
https://arxiv.org/html/2506.21263
https://www.telecoms.com/partner-content/china-unicom-leads-in-distributed-ai-training-with-breakthrough
https://www.fanaticalfuturist.com/2024/11/major-chinese-generative-ai-training-breakthrough-runs-across-distributed-datacenters/
@Dulaman My guess it that lack of compute will still be a constraint here, assuming the export restrictions on blackwell chips hold.
@mr_mino I agree with that as the default assumption. But my sense is that access to blackwell chips is key for centralised training runs. Doing decentralised training runs potentially allows China to overcome these restrictions and leverage the existing distributed compute overhang, outpacing Western companies.
epoch make a comment about these existing compute overhangs, comparing to bitcoin network:
(but this graph is about FLOP "per second" so not a great apples-to-apples comparison)
Claude tells me bitcoin network does about 1e32 rough-FLOP-equivalents per year:
@Dulaman Let's assume that China is running on H20s, which can do 1.5e14 FP8 FLOP/s, and which cost about $15K/chip. They want to complete a 1e27 FLOP decentralized training run lasting say 6 months, and can expect to have 40% MFU. This implies that they need 1.07 M chips, at a cost of $16B (but probably more like $30B due to costs of packaging, cooling, etc). Even if MFU is lower due to communication costs, this seems within reach if China is determined.
