TIL: Incorrect music metadata is a huge problem and is causing artists to lose millions in royalties. This disproportionaly affects independent musicians. But structured NLP can help!

explosion.ai/blog/love-wi...

Why share nature pics on Instagram. When research papers can do. 😄

"Memes are the product" says an analyst for openAI.

Interesting to see the text-to-software space grow. As models get better.

But, few things are hype though.

The text to software products are cool as models get better.

But I do wonder if falls into the bucket of no-code tools of the past. Were their are great for the first draft or basic software. But any precise customisation. There it becomes more difficult than actually coding the software yourself.

Skylar Payne's Cursor Guide is arguably one of the top practical AI development tutorials online. IMO

Sometimes the best thing you can do as a software engineer. A simple 10-line script or 30-second video clip.

Not anything fancy.

My experience with o1 and DeepSeek, are great for programming. But provide critical mistakes making the code un-usable.

Here the human always needs to be at the wheel.

LLMs are powerful tools, not magic solutions. They're part of your toolkit, not a replacement for human judgment.

Great essay by Jack Clark. (ImportAI) The new generation of thinking models. Mean curiosity is the biggest differentiator for users.

R1 showed can turn any open source model into a thinking model.

Things are about to get interesting.

"We going to stop using cars due to better gas mileage," said nobody ever.

This applies to the current compute demand debate right now.

Prediction: Massive Cost reductions coming due to DeepSeek.

The Flash Gemiai already has similar costs to Deepseek. Stay tuned for "mini" models getting a big upgrade.

Been seeing some takes - Post DeepSeek. About demand for compute.

No - Compute is not going away.

Much more efficient yes. But still lots of gains by scaling up new frontier models.

Tobi’s Substack | Tobi Olabode | Substack My personal Substack. Click to read Tobi’s Substack, by Tobi Olabode, a Substack publication. Launched 6 minutes ago.

Tobi’s Substack tobiolabode.substack.com?r=8netp&utm_...

Been looking write for longer form content - stay tuned.

I've tried DeepSeek - Great at Programming, Great at Writing and most importantly cheap.

AI Safety is done differently from Western AI companies. So writing is much more creative.

I won't be surprised if DeepSeek releases a big model within the next 3 months. Using the same arc as R1.

Think Compute will decline after DeepSeek? Not true.

It means you'll get more for your investment. The same billion-dollar spend on DeepSeek Arc enhances your capabilities beyond what was possible before.

But massive CapEx from Wall Street is over IMO.

Just my opinion, scaling a small, inexpensive model into a larger one while minimizing costs offers excellent value.

However, I don't see a strong incentive to reduce power consumption in this scenario. You just have more incentive for scale more.

This is already happening right now. GPT-4 intelligence as a JSON parser.

Expert investors asked 65.1% specific financial questions vs 50.8% from novices.

The real edge?

Knowing exactly what to ask your AI assistant.

Paper - AI, Investment Decisions, and Inequality

Everyone's shouting about 20% AI growth, but here's the inconvenient truth: transformative tech rarely creates overnight revolutions.

It creates compounding gains. Like electricity before it, AI's real power is in the slow reshape 🌱

Yep, I think it's further evidence for the internal "GPT-5/Claude 3.5 Opus" hypothesis.

The labs are keeping their frontier models private and using it to produce training data.

Seems like LLMs can deal with the cold start problem for new projects.

Fascinating pattern: Pro investors start broad, then drill down strategically.

Meanwhile, newer investors showed 21.5% query redundancy (vs pros' 14%). Experience shapes how we interact with AI. 🤔

Paper: AI, Investment Decisions, and Inequality

Sophisticated investors + AI = 18% better at predicting earnings

Everyone else + AI = only 7% improvement

Paper by MIT-Chicago Booth Reserach 2024

The American insurance market is not in a good spot - California wildfires to Florida hurricanes.

New pre-print! 🎙️🦉

Automated Note Annotation after Bioacoustic Classification: Unsupervised Clustering of Extracted Acoustic Features Improves Detection of a Cryptic Owl

papers.ssrn.com/sol3/papers....

#bioacoustics #machinelearning #conservation

LLM integration success = Base model + Prompt template + Data Strategy

Pro tip: Start with a pre-trained model like GPT-4o, then customise with your data and prompts.

Kevin Ko reports from the Pacific Palisades, where he speaks with one family who took the firefight into their own hands when the Palisades Fire closed in on... Pacific Palisades family fends off flames with their own firefighting equipment

A family saved their home with a custom spriker system.

https://www.youtube.com/watch?v=xs0DMUZ6_WQ