Fran Bartolić

Weather forecasts are ultimately used to make decisions, but most NWP and AI weather models predict the atmospheric state at discrete timesteps, while real-world decisions depend on decision-centric targets that are often nontrivial to recover from state snapshots. Our proposal is straightforward: learn the joint distribution over the atmospheric state and decision-centric targets directly, while evolving only the atmospheric state autoregressively. We illustrate the framework by modelling true daily temperature extremes despite using a 24h timestep for the state variables, and it generalizes to a broad class of decision-centric targets. We also show you can obtain near-SOTA state forecasts with off-the-shelf vision transformers on a lat/lon grid, with 1–2 orders of magnitude cheaper training and inference than competing approaches, including recent work from @GoogleDeepMind. Finally, we obtain stable S2S and long-range forecasts at daily resolution that smoothly converge to climatology, despite operating on a rectangular grid. Preprint now out on arXiv: arxiv.org/abs/2601.03753

Gemini 3 is the only frontier model that still regularly gets stuck in infinite token generation loops....

Google's @antigravity has become completely broken recently. Not that it every worked well though. The only reason I keep using it is because of cheap Opus 4.5 credits.

The main contributions of the paper are neatly summarized in Figure 1:

Weather forecasts are ultimately used to make decisions, but most NWP and AI weather models predict the atmospheric state at discrete timesteps, while real-world decisions depend on decision-centric targets that are often nontrivial to recover from state snapshots. Our proposal is straightforward: learn the joint distribution over the atmospheric state and decision-centric targets directly, while evolving only the atmospheric state autoregressively. We illustrate the framework by modelling true daily temperature extremes despite using a 24h timestep for the state variables, and it generalizes to a broad class of decision-centric targets. We also show you can obtain near-SOTA state forecasts with off-the-shelf vision transformers on a lat/lon grid, with 1–2 orders of magnitude cheaper training and inference than competing approaches, including recent work from @GoogleDeepMind. Finally, we obtain stable S2S and long-range forecasts at daily resolution that smoothly converge to climatology, despite operating on a rectangular grid. Preprint now out on arXiv: arxiv.org/abs/2601.03753

The reason people hate Zohran is because he’s a spoiled retarded communist. His parents are privileged, placeless global citizens and Zohran was able to float along without real jobs in NYC taking a swing at a ‘rap career’ while living in a Chelsea apartment owned by his mother. Despite this he plays the victim and cries about being persecuted and is a vector for every bad idea that afflicts downwardly mobile creative affluents. As his next act, he’s putting on, ‘Socialism: the Musical,’ with a bunch of other failed theater kids with a $100 billion annual budget in the US’s greatest city. It’s the stupidest timeline, he’s giving normal people all sorts of things to worry about (public safety, schools, etc.) that they shouldn’t have to, and it’s all in service of the grievances of a class of losers perpetually bitter that they can’t have the lifestyle in NYC that their income affords them.

Third pass through Melissa. GoPro in side window as different camera looking forward shooting in ultra high res 8k. Not sure when that might get processed as the file turned out ridiculous. Barely had HD space for it and MacBook Pro promptly chocked when I tried to edit it

@EdsenTheWeather @GoogleDeepMind @ViktorCikojevic @LevangSam @PauliusRauba Thanks! We're hoping to publish something in a month or so.

@fbartolic @GoogleDeepMind @ViktorCikojevic @LevangSam @PauliusRauba Hi! This is really amazing work. Congratulations to the team! Do you have an update on the ETA for the arXiv paper/benchmark comparison?

I’m excited to share @salientpredict’s newest generative AI weather model, GemAI v2. It delivers 200-member ensemble forecasts with a 126-day horizon and beats @ECMWF’s IFS ENS by a substantial margin across multiple variables and timescales.

@WeatherMatrix @burgwx Here's a similar animation showing the probability of maximum daily wind gust (rather than mean surface level winds) above a 98th historical percentile threshold.

@WeatherMatrix @burgwx Salient forecasts aren't publicly available but here's a short animation I've made showing the probability of extreme winds from today's GEFS and GemAI forecasts.

I usually don't post about a new feature before it's complete, but in the interest of timeliness with an uncertain path for invest 98L, I've been working on implementing a 360-member grand ensemble consisting of 9 ensemble suites (NWP + AI). Here is its latest projection for 98L:

@benln Where is it? I'm on the waitlist :/

Cafe Cursor NYC is open

@aakaran31 I'm curious, why did you use M-H instead of a gradient-based sampler such as HMC?

We employ a Metropolis-Hastings (MCMC) approximate sampler, which iteratively updates a generation by partially resampling new candidates, accepting with probability depending on pᵃ. To make this approach suitable for LLMs, our algorithm integrates Metropolis-Hastings into autoregressive generation, building up to a sample from pᵃ block-by-block. (6/n)

We found a new way to get language models to reason. 🤯 No RL, no training, no verifiers, no prompting. ❌ With better sampling, base models can achieve single-shot reasoning on par with (or better than!) GRPO while avoiding its characteristic loss in generation diversity.

@GoogleDeepMind This was a long time in the making. Shoutout to @ViktorCikojevic, @LevangSam, @PauliusRauba, and everyone else who contributed to this release!

There’s a lot more to say about GemAI v2. The model is already operational and available through our GUI and API. We are planning to submit a paper on arXiv soon, along with a benchmark comparison against @GoogleDeepMind’s latest GenCast and FGN models