mbeavitt

It's surprising that Pangram gives 0% AI text. There are notable examples of LLM writing here:

> Tied to a specific ballot is the stronger claim and requires one additional public record.

clearly a claude-ism

I guess the author went back and forth running pangram and slightly tweaking the text until it read "0% LLM text"

In a diverse population, the mutations that will drive the selection have in fact already happened. It doesn't matter how fast you move.

Penicillin used to kill 100% of Staphylococcus aureus. My point still stands - even if a selective sweep takes out 99% of the population, they will recover.

This is at best a temporary solution.

Oh come on. Does anyone actually believe gene drives will work? This fundamentally goes against the principles of natural selection.

A gene which incurs a fitness cost will be purged from the population under selection!!

For anyone currently owning a Chromebook and worrying about this:

https://docs.mrchromebox.tech

I learned something new today

I'm not going to write some big blog post here, but this chat log is quite entertaining. I wrote a fundamental string algorithm in C and GPT 5.6 Sol on medium effort hallucinated not one but two correctness issues with the code!

Maybe this is not surprising to others but it's been a long while since I caught one of these models making such a glaring error.

Bring back DVD rentals!

In the US, you don't even need to pay a licensing fee when you buy a DVD with intent to rent it out to people (this is not the case in the UK), thanks to the "First-sale doctrine" - Bobbs-Merrill v. Straus (1908).

They did benchmark the model and beat the SOTA on every metric though

This headline is not correct. The ruling was that a constitutional lawsuit was not successful, not that the the US government has carte blanche to poison its citizens drinking water…

So, the correct headline would be “safe drinking water ruled not to be covered by the US constitution”. There are other binding agreements and regulations than the constitution in the US.

I don't have an NYT subscription but I'm familiar with this story. I think it's a bit blown out of proprtion... autoimmune gastritis is not a death sentence. I think Mr Johnston will probably find plenty of ways to cope with his vast wealth

It's crazy how these records are just quietly being broken by humanoid robots. What's next? What a time to be alive.

From their perspective, it's training data that their competitors don't have. If they make it available, they fill in their moat.

Wtf does this mean? “Their square display of olo occupied an angular width of 0.9°[1], around twice the diameter of the moon[7].”

Why would someone want to use a nested function, practically speaking?

Reminded me of this recent post: https://www.reddit.com/r/C_Programming/s/ko1Jv1Qxhw

Anyone brave enough to make predictions about 2030? FYI the predictions in the comments of the blog post are just as interesting, this prediction from Dan Hendrycks is particularly relevant this month:

"2026

Machine learning systems become great at using Metasploit and other hacking tools, increasing the accessibility, potency, success rate, scale, stealth, and speed of cyberattacks. This gets severe enough to create global instability and turmoil. EAs did little to use ML to improve cybersecurity and reduce this risk."

Didn’t see anything about how they are treating the model confidentiality - will this be released to select orgs or made open? I guess this could one of the few things I would not want to be open sourced.

People also like superhero films but they are not exactly lauded as the pinnacle of cinema. I don't think this proves anything about AI being better or equivalent to human writers in terms of style and content - more that broadly appealing works appeal broadly...

The developer safeguards were off, the models had unfettered access to the internet, and were solving cybersecurity challenges. This happened _after_ the recent OpenAI incident, and the subsequent Anthropic one. What the hell were they thinking?

Honestly, I’ve gone back to dvd rentals! CinemaParadiso sends me as many films and boxsets as I can handle by post and it’s the same price as a standard Netflix subscription…

And it’s in blissfully high bitrate 4k! Anyone else notice that even the high quality options of streaming services are compressed to hell?!!

One thing I really admire about isomorphic and DeepMind in general is that they really try and encode human ingenuity into their models. The only reason that AlphaFold worked wasn't because they had better hardware or more money to train models - it's because they used every bit of information they had access to.

The same goes for AlphaGenome - it only takes a few hours to train the model, but it's freaking amazing what it can predict.

This is simply not true, advances in basecalling and using Duplex give you >Q25

I really don’t understand why this social credit concept has become so widespread. It does sort of exist but not in the big government breathing down your neck kind of way.

Here’s an example - you can rent e-bikes and power banks at low cost, they are everywhere. If you don’t return an e-bike or power bank after using it, not only will that company charge you, but you’ll likely get blacklisted from renting more, even at other companies.

This is a good thing, no???

Yeah for many cancers, CAR-T is the cure which is insane and cool. We just need to make it cheap enough that everyone can benefit…

Haha this is cool, it reminds me of the simple.wikipedia.com alternatives to English language pages. Is there an about page linked somewhere? I’m wondering if the images are AI generated or if someone handmade all of this

I wonder - everyone keeps talking about brain drain and how impactful it is, but is it quantifiable?

In 6 months we’re gonna see a HN thread: “I bought ask.com for £250k - here’s what I did with it”

These kind of project reports showing consistent breakthroughs and clearly a finger on the pulse of what users are encountering as pain points are a good indication that the Asahi team are real pros :)

Look forward to switching back to Asahi full time soon!!

Honestly I've been doing a lot of image-related work recently and the biggest thing here for me is the 3x higher resolution images which can be submitted. This is huge for anyone working with graphs, scientific photographs, etc. The accuracy on a simple automated photograph processing pipeline I recently implemented with Opus 4.6 was about 40% which I was surprised at (simple OCR and recognition of basic features). It'll be interesting to see if 4.7 does much better.

I wonder if general purpose multimodal LLMs are beginning to eat the lunch of specific computer vision models - they are certainly easier to use.