dexta.bsky.social

So many people naively buying into Mythos hype. When it releases (+ spud, Gemini 3.5 etc) to much less hype, there will once again be zero reckoning or apology, and the next cycle will begin

Seeing health as a virtue is not an attitude any health professional of any kind should have.

I have cancer. mRNA technology is already being used to develop some of the next cancer cures. He’s condemning some current and many, many, many future cancer patients to early, preventable deaths.

This is getting at the fundamental problem with EA, Rationalism etc. They hugely overemphasize the present, which is obviously ironic for long-termism and similar sub-identities of these things people take on.

An "internal model at OpenAI" has solved 5 Erdos problems, all reported in the paper below. All of those have now been listed as Category 1 = Fully AI-generated solutions here: github.com/teorth/erdos...

Mass hysteria

OpenAI will release a 'Mythos-class' model in the next few months, e.g. spud, followed by the other labs. Not much will change and all these posters expecting a cyberpunk fantasy to play out will move on to the next wave without blinking

So many people naively buying into Mythos hype. When it releases (+ spud, Gemini 3.5 etc) to much less hype, there will once again be zero reckoning or apology, and the next cycle will begin

It's kinda wild that the one thing Anthropic claims to focus on, coding/agentic coding, OpenAI beats them at whilst multi-tasking What Anthropic are actually the best at is marketing. There's an army of people using Claude Code who've never even tried other models or harnesses. Genuinely impressive

There should be more data centers

Is Gemini 2.5 PRO Preview 03-25 still the most powerful AI model ever released? I'm genuinely only 90% sure it isn't

AI bros' version of attack helicopter

Some people really do get envious and territorial about activism

One of the words below was AI generated, can you tell which? 1. green 2. low

When a pastor was asked if he's excited for super intelligent AI, he said of course, as being super intelligent it would quickly realise Jesus is lord and become a super Christian.

People posting identical side-by-sides of minebench performance and acting like there's a huge difference

1. Never assume something isn't in the training data. Search the data. 2. Never assume the model is the same as before unless you have access to the weights 3. If you use some else's API, your benchmark isn't private

The most details I've seen yet on exactly how the military uses Claude. Really great reporting by @taracopp.bsky.social @lizzalichka.bsky.social and Ian Duncan wapo.st/4b15X9p

The White House just posted a video mixing real footage from the Iran strikes with a killstreak animation from Call of Duty: Modern Warfare 3 x.com/WhiteHouse/s...

People really need to call the AI bubble the data centre bubble. It's the financing behind the AI data centres that looks precarious, not the use of the models themselves.

New York Times is just a platform protected by Section 230

I literally switch between Claude Code, OpenCode and Codex everyday. It's almost effortless...

This still explains most of the AI posters on this site

New paper on a long-shot I've been obsessed with for a year: How much are AI reasoning gains confounded by expanding the training corpus 10000x? How much LLM performance is down to "shallow" generalisation (approximate pattern-matching to highly-related training data)? t.co/CH2vP0Y7OF

Thinking of starting a conspiracy theory that Epstein was actually King Arthur returned

Was discussing something like this with a friend last night, but it is fascinating how many parents now and forever take a weird pride in not knowing the basic details of the shows their children enjoy.

The only financial problem major AI companies have is no moat, which *is* existential. Massive debt, negative gross margin and copyright violations are all tried and true strategies, not problems. Most of the upcoming data center capacity isn't even needed, so a crash won't affect them much either.

If he didn't talk much, people would assume he was an eclectic genius

A whole thread of people mistaking being British for being mentally ill

Every time the make a new prediction, their old predictions should be listed first for all to see.

You only get to say we aren't ready for something if you can describe how to be ready for it, otherwise how do you even know we aren't ready? Vague fear posting.

also IF the AI doomerism was correct and we are in the process of building literal Skynet the ONLY ethical way forward would be to deploy lethal military violence against the AI industry and it's researchers worldwide. If you aren't suggesting that, you don't believe this is end-the-world serious.

Unfortunately the coding language I am currently using only has thousands of lifetimes of data on the internet, so LLMs just aren't very good at it.

Epstein will be a trigger word for them, but if you described his actual crimes without the name and fame they would barely have a problem with it. They don't care, and if Epstein was alive he could easily grift his way to influence like Tate and all the others.

this feels like first-time.gif for some people for a new frontier model, so the usual: - blog posts are marketing. be skeptical - benchmarks mean less than how it works for *you* on *your* work - things that lower token spend are good; things that increase it should make you suspect (swarms)

It's very cringe to see people falling for the theatre of moltbook. None of them check or confirm anything. Jingling keys taken at face value yet again.

i would trade a kidney for industry critics who were attempting to be factually correct