Emile Jansons

[Ethan Mollick @emollick]

I've had access to Fable for a bit. A genuine jump in capability, I could feed it a 15 page design document for a project and it would work for 9+ hours and deliver terrific results.

But working with it is weird & weirder is coming

Lots of examples: https://t.co/HptkYunBzr

[Emile Jansons @emile_jansons]

@Daniel__78 @BrianDoher37387 @skdh A good test for the ice lofting hypothesis is to compare transit freq. in the nuclear event -day to the +day. Cross references with major volcanic eruptions could also help test the high altitude ice crystals explanation.

[Space Jellyfish Alerts @spacejellyalert]

Space jellyfish alert!

Starlink 17-14 from SLC-4E in California on April 22, 2026 at 7:56:39 PM PDT could produce a space jellyfish. This occurs when a rocket and its expanding exhaust plumes are illuminated by the sun and observers are in local darkness.

The rocket will enter sunlight 01m25s after launch.

See a detailed prediction for your location: https://t.co/KIadNi37so

Information as of 2026-04-23 01:08 UTC

[Yohan @yohaniddawela]

2.6 million flood events were hiding in plain sight inside the news. Google just turned them into a training set.

Satellites and traditional databases are too slow to catch urban flash floods. The United Nations disaster database holds about 10,000 high-impact events. You can't train a global AI model on 10,000 data points. Accurately modelling climate resilience requires massive historical baselines that simply didn't exist.

Google researchers built a pipeline called Groundsource to fix this data desert. They pointed Gemini at decades of unstructured global news reports across 80 languages. The LLM translated the text and ran a strict verification process. It separated actual past floods from future warnings. It anchored relative dates to publication timestamps. Then it extracted granular street-level locations and mapped them to standardised spatial polygons.

They essentially turned unstructured text into a structured geospatial dataset. Groundsource generated 2.6 million verified flood events spanning 150 countries. Manual reviews showed 82% of these extracted events were accurate enough for real-world spatial analysis.

This historical baseline immediately upgrades urban flood forecasting. Google is now rolling out near-global urban flash flood forecasts up to 24 hours in advance.

Physical sensor networks take decades to build and deploy. Advanced data engineering provides a massive shortcut.

The world's local news is a perfectly valid historical sensor network...

[Will Manidis @WillManidis]

https://t.co/bDCdltFo0A

[John Carmack @ID_AA_Carmack]

This is good advice.

[RAW EGG NATIONALIST @Babygravy9]

The full depth of liberal Australian insanity on display here. Giving one of the most important archaic human fossils ever discovered to a group of aborigines so they can bury it somewhere it will never be found again. The people responsible should be put in trial.

[Ethan Mollick @emollick]

In a large representative sample of humans compared to GPT-4: "the creative ideas produced by AI chatbots are rated more creative [by humans ]than those created by humans... Augmenting humans with AI improves human creativity, albeit not as much as ideas created by ChatGPT alone”

[Rohan Paul @rohanpaul_ai]

Your brain's next 5 seconds, predicted by AI

Transformer predicts brain activity patterns 5 seconds into future using just 21 seconds of fMRI data

Achieves 0.997 correlation using modified time-series Transformer architecture

-----

🧠 Original Problem:

Predicting future brain states from fMRI data remains challenging, especially for patients who can't undergo long scanning sessions. Current methods require extensive scan times and lack accuracy in short-term predictions.

-----

🔬 Solution in this Paper:

→ The paper introduces a modified time series Transformer with 4 encoder and 4 decoder layers, each containing 8 attention heads

→ The model takes a 30-timepoint window covering 379 brain regions as input and predicts the next brain state

→ Training uses Human Connectome Project data from 1003 healthy adults, with preprocessing including spatial smoothing and bandpass filtering

→ Unlike traditional approaches, this model omits look-ahead masking, simplifying prediction for single future timepoints

-----

🎯 Key Insights:

→ Temporal dependencies in brain states can be effectively captured using self-attention mechanisms

→ Short input sequences (21.6s) suffice for accurate predictions

→ Error accumulation follows a Markov chain pattern in longer predictions

→ The model preserves functional connectivity patterns matching known brain organization

-----

📊 Results:

→ Single timepoint prediction achieves MSE of 0.0013

→ Accurate predictions up to 5.04 seconds with correlation >0.85

→ First 7 predicted timepoints maintain high accuracy

→ Outperforms BrainLM with 20-timepoint MSE of 0.26 vs 0.568

[the tiny corp @__tinygrad__]

We switched from AMD's driver to our AM driver...and now 4 GPU Llama is faster on red box than green box!

[Ethan Mollick @emollick]

Reminder for the new semester

When researchers secretly added AI-created papers to the exam pool: “We found that 94% of our AI submissions were undetected. The grades awarded to our AI submissions were on average half a grade boundary higher than that achieved by real students.”

[Ethan Mollick @emollick]

For a couple years, I have been using "otter using wifi on an airplane" as my test for the improving ability of AI image and video systems.

But this is now too easy a challenge for current AI, so what if we kicked it up a notch (these are all veo 2 clips).

[Deedy @deedydas]

The Superhuman AI for Poker is a timeless research paper.

Poker, unlike Go/Chess, is an imperfect info game. Pluribus, made by now OpenAI engineer, beat 5 pros decisively in 10,000 poker hands, winning 48 milli big blinds/game.

Made a mini demo on rock-paper-scissors too:

1/5

[Jeremy Howard @jeremyphoward]

The tradition of folks that do not have the expertise to use notebooks effectively then going on to blame the tool continues for another year.

Bring on 2025!

[htmx.org / Viceroy of Corn Dog (same thing) @htmx_org]

the first three chapters of our book, https://t.co/KWTfVdQtvD (free online) cover the conceptual basics of web development 1.0

the perfect gift for the zoomer in your life who doesn't know what a form tag can do

zoomers & embittered, washed up gen x devs, stronger together!

[Jeremy Howard @jeremyphoward]

I'll get straight to the point.

We trained 2 new models. Like BERT, but modern. ModernBERT.

Not some hypey GenAI thing, but a proper workhorse model, for retrieval, classification, etc. Real practical stuff.

It's much faster, more accurate, longer context, and more useful. 🧵

[htmx.org / Viceroy of Corn Dog (same thing) @htmx_org]

htmx is so efficient you can literally shitpost on twitter 23 hours a day and still ship software when you use it i'm not even lying