Based on rumors that scaling is fitting diminishing returns, this means that US companies may delay releases for longer and Chinese companies may be able to turn their compute limitation into a temporary advantage. Slightly increasing previous confidence, but it is a rumor, so not by much.
No Scores Yet
Relative Brier Score
0
Forecasts
0
Upvotes
Forecasting Calendar
Past Week | Past Month | Past Year | This Season | All Time | |
---|---|---|---|---|---|
Forecasts | 1 | 4 | 7 | 7 | 7 |
Comments | 1 | 3 | 3 | 3 | 3 |
Questions Forecasted | 1 | 2 | 3 | 3 | 3 |
Upvotes on Comments By This User | 0 | 0 | 0 | 0 | 0 |
Definitions |
Active Forecaster
Why do you think you're right?
Why might you be wrong?
Rumor is false, scaling continues to work. Or the timing just doesn't work out.
Why do you think you're right?
Multiple Chinese labs have demonstrated the ability to enter top levels, including o1, Alibaba, and Deepseek. But:
* OpenAI, Anthropic, and Google exist.
* Top AI labs have multiple entries in LMSYS, for example currently OpenAI occupies slots 1, 2, and 3. This means that the leading company will likely occupy multiple top slots, making it hard for second rate models to edge in.
* Chinese labs are playing catch up and have to focus on computational efficiency due to shortage of hardware.
By rights, this should mean that the probability is very low, between 0-5%.
However:
* AI labs are not collaborating and can therefore variance is increased (therefore, increase the base rate somewhat).
* Top AI labs are reluctant to release new models because any subpar model will affect fundraising. As a result there are lulls where no new model is released, leaving opportunities for outsider models to take top slots beyond their actual capabilities. Moreover, other labs can specifically target these lulls to gain publicity and funding (~20% this scenario occurs)
Why might you be wrong?
Most likely scenario is one of the top labs releases an extremely strong model that cannot be beaten for an extended period of time. Especially if there are limits of the efficient approach and pure compute remains the bottleneck, it should be difficult to get to the top 3 even using a hunting the lulls approach.
Active Forecaster
Continue previous forecast.
Active Forecaster
Firstly, the 6 month market has a percentage of 54% while this one only has 52%. The likelihood of war in 12 months should be higher, and the 6 month market has more forecasters.
Secondly, Israel has been increasingly aggressive against Hezbollah, to the extent that they are functionally at war without declaring it. The most likely scenario where this resolves NO is that Israel remains de facto at war without declaring it. The likeliest scenario for war occurring is that Hezbollah feels forced to declare war, at which point Israel can be justified in whatever it wishes, including ground invasion. To the extent that Hezbollah does not believe it can gain anything by this, it may simply try to endure instead.
Scenarios with No War:
* Israel functionally defeats Hezbollah through espionage and targeted strikes: 1-5%
* Israel suppresses Hezbollah through such means but leaves them a path to survival: 15-35%
* Israel is pressured to stop aggression by Western powers: 5-10%
* Iran develops nuclear weapons: <1%
* Nothing ever happens (prediction market meta prior): 1-10%
Scenarios with War:
* Israel attacked by coalition of forces, including Hezbollah and Iran: 1-5%
* Hezbollah feels forced to declare war on Israel: 25-50%
* Israel declares war on Hezbollah to achieve specific strategic objectives: 5-15%
* Israel defeats Hamas quickly and feels emboldened to take on the next challenger: 1-5%
These numbers are all made up and not based on anything
Why do you think you're right?
Updating based on continued rumors that current scaling laws have hit a wall, and rumors of DeepSeek R1 as described in https://buttondown.com/ainews/archive/ainews-deepseek-r1-claims-to-beat-o1-preview-and/
Why might you be wrong?
DeepSeek is exaggerating performance, or other dark horse candidates unexpectedly release a model