A Random Thought

Forecasts for 2025

On the Lightcone 2025 Forecasts

I recently saw this episode of the YC Lightcone podcast, where each of the hosts makes a forecast for 2025. I wanted to comment on the predictions because I think they are really interesting and open up entire discussions on AI, crypto and where we think those fields will go in the near term.

TLDR: the forecasts and my opinion on their likelihood

  1. Diana Hu: AI research to win another Nobel Prize
    • Very Unlikely in 2025, but very fun
  2. Harj Taggar: Crypto goes mainstream (via stablecoins)
    • Likely for cross-border payments,
    • Very Unlikely in the US for direct payments,
    • Not Gonna Happen globally for direct payments
  3. Garry Tan: The fate of DOGE is linked to the fate of DOGE
    • Likely, but hard to measure
  4. Jared Friedman: Zoom call with an AI
    • Very Likely, with scale caveats

A quick aside: A scale for "verbal" forecasting

We often make probability judgements in uncertain terms. It's helpful to define a scale for our assessments, so that the communication is clear. I'll use the following wording for my forecasting and likelihood ratings:

I know this scale is all over the place with interval sizes, but I wanted a nice 7-point scale. It also assigns higher resolution to the ends of the scale, so that I can better point-out how certain I am about stuff.

Diving into each forecast

AI Research will win another Nobel Prize

This is the most interesting forecast, in my opinion. While I think it's unlikely to happen in 2025, in an extended horizon its almost certain.

The 2024 prize in physics highlighted foundational research in the development of neural networks. Hopfield and Hinton laid much of the groundwork for NN models, and Hinton later collaborated on many of the advances that brought NNs into the mainstream. Meanwhile, the prize in chemistry recognized advances in modeling and predicting protein structure, which has large research impacts in biological and pharmaceutical research. Here, the DeepMind duo were recognized much more directly for creating a model with great performance and impact.

While there is much more work to be recognized, I'd wager that research in computer vision (DNNs, CNNs, segmentation) has the greatest chance of earning the next Nobel for the field due to its impacts in Physics and Medicine. Vision models have been used with great effect to analyze both astronomical and medical data, leading to gains in scale and automation that would be virtually impossible otherwise.

I find it very unlikely that the committee would award the Nobel in Physics to "AI research", in broad terms, for two consecutive years: that hasn't really happened in recent years, and the prizes tend to favor older, proven research. For the prize in Medicine, I honestly have no idea what I'm doing, but the prizes do seem to favor more widely adopted, direct impact research. Overall, I'm going with a Very Unlikely rating, but would be happy to be surprised.

What about Math1?

Diana also mentions the significant work being done in AI for Math research. This is the most intriguing avenue for me at the moment. Systems that are capable of reasoning about logic and theorem-proving at a high level will revolutionize how we think about AI and maybe even increase our own formalism.

Reasoning models (like o1 and o3) seem capable of treating more complex mathematical problems, which maybe will allow large-scale exploration of hypotheses - like having a legion of people test your ideas at a low cost. At the same time, proof-writing models (eg. AlphaProof) are already able to solve IMO problems using theorem-proving systems, and maybe will empower mathematicians to explore new proofs, or validate proofs faster - which itself would propel the field forward.

I'm certain these lines of work will lead to great research in the future. An Abel Prize will most likely come, though I'd say that's not gonna happen in 2025.

Crypto will become mainstream

Harj's prediction is that by the end of 2025 most people will have used crypto to buy something, "a coffee or a book".

As a gut reaction, I'd say that that's just not gonna happen, but that's me thinking as someone living in Brasil, working in Brasil. Brasil has free, instant, universal payments via PIX. For some other countries, and especially for people who move money across borders, I see a lot of potential in stablecoins and instant payment solutions.

The Fiat-USDC-Fiat Pipeline is Real

The argument for cross-border payments is solid: transactions in stablecoins typically involve low fees, and the same network can be used to move money effortlessly to anyone, anywhere. If there is a good solution to bridge fiat-stable and stable-fiat in many destinations at a low cost, I can see these applications taking market share from traditional players like Wise.

All in all, this seems likely to go into the mainstream, though I'd be surprised if any of these players became dominant in the space.

Direct Payments are Hard

I find it very unlikely that crypto-payments will take much of the market in the coming year.

The argument to me is much less solid here. With the adoption of FedNow making instant payments easier to implement, I'd expect credit and debit transactions to take even more market share in the short term. It's not that hard to imagine a cheap instant transfer solution coming in the near future.

The main reasons for a payments solution to quickly gain market share are lower costs to consumers or companies, and being easily accessible. In that sense, if traditional systems are able to provide cheap transactions that use the existing payment methods (checking accounts, cards), there is little reason to adopt crypto-based payment methods.

Almost all adults in the U.S. have a credit or debit card and access to a checking account. Besides that, moving money in-and-out of stablecoins typically goes through an existing payments layer, making these payment methods inherently harder to access. The main reason for change would have to be lower costs, but conventional solutions can achieve very low operational costs for instant payments, as seen in India, Brasil, and many other countries.

If DOGE works, DOGE will go up

This forecast will be much harder to break down and measure. Garry proposes a mechanism: if DOGE is successful, spending will go down; this will cause interest rates to go down, which will bring crypto prices up.

The problem here is how to assert the "if DOGE works" part: what will we use to measure the success of Musk and Vivek's department?

None of these seem to capture the full story, and I'd wager it'll be difficult to have a simple response by the end of the year. Nevertheless, I find it likely that the department will pass some proposals in the first year, and they'll likely find the most popular gains early on. If they bring significant spending cuts that are popular with the general public, DOGE will go up.

One other thing that'll be hard to assess is whether the variations in Dogecoin will be caused by actual rate cuts, by the expectation of rate cuts, or by the overall likelihood of positive crypto regulations in the Trump presidency.

My own prediction here is that DOGE is likely to live through the first year, and that Dogecoin is likely to go up (not financial advice, and I don't hold any). Whether there'll be a causal relation there, I honestly don't have a clue.

Zoom call with an AI

Someone, somewhere, will most likely post a demo just like this: two people talking, just chatting about travel and sports and the true meaning of being alive. All is very fine, except that one of them is an AI.

Video generation and syncing has advanced very quickly: a couple years ago AI couldn't make a decent scene, by the start of 2024 it could but had some obvious giveaways, now it looks like a weirdly produced movie. By the end of the year, I think it's very likely that one of the startups working on digital characters will put together an agent that crosses the uncanny valley and feels like talking to a real person.

There is a large caveat: this will require large amounts of compute to run live, and will certainly not be comercially available at any scale. Not by the end of 2025, at least. That in no way invalidates the achievement and the applications.

Criticizing is easy. Here are my (late) predictions.

I realize I'm a bit late to this, as it's already 2025 by a lot, but I want to end by making my own bold predictions for the coming year. I've organized them by how much I like them.

  1. LLM performance in coding and reasoning tasks will stabilize, as measured by popular benchmarks
    1. I'm sure that coding systems will get more and more capable, but we seem to be stabilizing in terms of performance per compute.
    2. Throwing compute at the problem isn't really viable for commercial use-cases, so commercial systems will likely stabilize and evolve more in line with the advances in compute and drops in cost.
  2. Apple Vision Pro cancelled
    1. This will be an easy one, but it's relevant I guess.
  3. Bird Flu will continue to rise in the US, but person-to-person spread won't be confirmed by health agencies (CDC, WHO, etc)
    1. Honestly, this is more hope than prediction.
    2. If I'm wrong, please don't bring back Zoom Happy Hours.
  4. A large (> 8h) airline outage is caused by ransomware (in the US or Brasil)
  5. Russia and Ukraine will still be at war by the end of the year
    1. The US will reduce military aid to Ukraine.
    2. The EU will step up their aid in return, but will not be able to maintain the same level of support.

Footnotes

  1. Some YouTube comments hastily pointed out that there is no Nobel Prize in Mathematics. The Abel Prize, however, is an essentially equivalent prize established in 2003. In the same sense, the prize in Economics is not an original Nobel Prize, but was established much later to recognize great contributions in this field.↩

#comments