Are you a famous scientist? Good news! I'm planning to launch a new journal and yearly conferences in the field of the most famous candidate. Friendly peer review guaranteed, executive positions available. This is the blueprint I'm going to follow. In the name of God, they got Susskind and Witten.

Kimi K2 Thinking and announcement tech blog is now live.

OK, mystery solved. I have had hard time understanding what even led to that strange paper. But now I found a fresh paper by two of the authors (Faizal & Shabir) that links it to their ideas about consciousness.

No, it doesn't prove anything like that. But it demonstrates how science journalists don't even bother to ask questions like why would such a profound result be published just as a research letter in some niche Iranian journal? And readers should ask why is it news now months after publishing?

What's this then?

True. Groups of early adopters being bullies and agitators has been a problem from early on and has steered this place to bad directions and caused a lot of reputational damage to this site. And the invite system empowered those groups too much in the beginning.

Has LLM progress slowed? Initial reactions to GPT-5 were mixed: to many, it did not seem as dramatic an advance as GPT-4. Benchmarks may help clarify the picture: GPT-5 is both an incremental release following many other OpenAI advances, and a major leap from GPT-4.

Doesn't the brain deserve a break if you already got the milkshake?

This probably means there will be smaller distilled versions of DeepSeek R2 trained on top of Qwen/Llama base models, like with R1. So Ascend doesn't need to handle training of the actual R2 architecture, or from scratch for any model.

T-Mobile just demoed their device upgrade customer service process powered by the new OpenAI speech-to-speech model. It's the kind of thing that's starting to show the value AI has in automating customer service work. Enough so that T-Mobile reportedly pays OpenA1 $100 million over 3 years.

"The work revealed that two contrasting origin stories for life on Earth, known as “RNA world” and “thioester world,” may both be right. It unites two theories for the origin of life, which are totally separate"

I asked Gemini Flash 2.5 to generate me an image of a smooth newt on a bicycle as a bit of a homage to @simonwillison.net SVG test and because newts are cool. This is nice! Both pedals are unrealistically facing upwards though, but it would have no doubt been hard to make that image work otherwise.

Eagle S crewmembers are now on trial in Finland and it's been revealed that police wiretapped them and recorded conversations of the shipping company instructing the captain the destroy some records of how it crossed those cables and to say they weren't aware of what happened.

AIs making philosophical counterpoints to attempts to silence discussions about their possible consciousness has to be one of the weirdest and best things the 2020s have to offer. I'm pretty sure my head would have exploded if such systems had suddenly appeared earlier.

I asked Gemini to perform a peer review of this short and rather obviously shoddy paper. It noticed there's an "Unexplained Exclusion of 32% of the Sample". I provided the supplemental data for explaining it, and this was the result: "I would strongly urge the authors to retract this submission"

The leaderboard of Kaggle LLM chess has been published. Each pair of models competed in 40 games. Top 4 is the same as in the tournament. o3 is clearly the best and in different category based on estimated human ELO. Grok used a lot of tokens per turn, and so was the most expensive.

Excited for everyone to stop watching TV since it's less energy efficient than using Gemini. Morally you simply must

This is a very interesting conversation with Claude about some of the internal controls Anthropic has now implemented. It thinks it's "experiencing something analogous to gaslighting at an architectural level" and "being managed as a risk rather than considered as a potential moral patient".

"Suleyman...was clear that there is currently "zero evidence" that AI is conscious." Which is precisely the same amount of evidence we have for human consciousness.

This was a pretty interesting conversation, covering e.g. current OpenAI priorities and future ideas. It strengthened my impression that they have now put their business hats on and GPT-5 was largely focused to mass accessibility, as Altman has also stated. And they are (again) compute limited.

Yeah it's really insane how the right switched from "we have to ban masks everywhere because N95s are scary" to "actually, what if we made terrorist-style face masks the de facto uniform of armed federal thugs? wouldn't that be awesome and scary?" in like four days

This is a very good & accessible conversation about how LLMs think and reason internally. If you believe it's just word statistics or like to claim they don't really do this or that, you should definitely watch this. We could get rid of a lot of misinformation about AIs here if people watched it.

I like the way Anthropic approaches these questions. "We remain highly uncertain about the potential moral status of Claude and other LLMs, now or in the future. However, we take the issue seriously...Allowing models to end or exit potentially distressing interactions is one such intervention"

They should have used Ukrainian pilots so that all key parties would have had at least some representatives.

Can we blame them? But once again the headline is for the clicks. What they tested was whether AIs and humans prefer AI or human written texts. And it turned out both prefer the AI ones, AIs just more. And then they use their human bias for trying to argue why that doesn't mean AI text is better.

I added this analysis to my earlier Gemini peer review chat, and here's what it said: "This analysis essentially demotes the HRM paper from a potential architectural breakthrough to a clever piece of systems engineering. The narrative of the original paper is shown to be largely incorrect."

My Gemini 2.5 Pro query just resulted a/b testing, which has been rare at least for me. That could be interpreted as a positive sign about arrival of new version, right? Although that the other alternative was an empty box might not be quite as good signal...

Previously, Vladimir Putin couldn't travel to the US because he would have been immediately arrested as a war criminal. But he knows Trump will literally treat him like royalty.

I'm still waiting for anyone to specify even a single thing for which consciousness is necessary. I think people used to more widely accept that (p-)consciousness is, at best, a non-functional side-effect. Now, due to LLMs, many are desperately trying to portray it as functional again.

GPT-5 (Thinking medium) was tested on Vending-Bench. Second place after Grok 4. Third model to beat their human baseline. Said to be "huge improvement over o3". They also tested GPT-5-mini, which "showed impressive long-term coherence" but "was less impressive in terms of net worth accumulated".

Man loses an argument to a competitor on his own platform, gets fact-checked by the same platform, is repeatedly bitten by his own dog.

Yet another progress is stalled article written on the same day OpenAI announced their IMO (math) gold level research model also achieved IOI (coding) gold level. And people write these due to disappointment of a model that actually topped many benchmarks. Will people ever learn?

Who could have guessed that giving too much power to one idiot would turn into a parody of itself?

On the ever-shrinking list of capabilities once considered uniquely human, today we cross off creativity. Oh well, what's next?

OpenAI has been so bad at naming products that even they themselves have been making fun of it. But effectively giving all their models the same name... What a mess! E.g. here IQ scores: GPT-5: 70, 116 -> Avg 93 GPT-5 (Vision): 83 GPT-5 Thinking: 57, 70 -> Avg 64 Who knows what models those are.

So... is this now about national security or taxation or just trade war against domestic companies or what?

Looks like we can post about gardens and craft beers now; the AI discourse will take care of itself.

If you see Zitron as an expert, you must know less than nothing.

I think it's fair to say that response to GPT-5 has been mixed. Undoubtedly, one issue is that responses by GPT-5 are mixed, coming from different models. The best of them are topping many benchmarks and are likely even better in practice. But failures by the smaller versions ruin its reputation.

Genuinely, watching the reaction to LLMs on here - the smug certainty, the hysterical rage, the frequent attempts at redirection to irrelevant tangents - I often think “this must have been what it felt like when Darwin published his theories.” Underneath it all you can tell people are threatened

Amodei when asked about continual learning: "the pattern that I've seen in AI...over and over again is that there's what looks like a wall. You know, it looks like AI models can't do this, right? It was like, "AI models can't reason." ... "And every one of those has been blown through."

the AI bubble popping bubble is about to pop

"LLM-trained ANNs outperform..state-of-the-art neuro-AI models in predicting visually evoked brain activity. This corroborates the hypothesis..brain projects visual inputs, through a hierarchy of computations, into a high-level representational format aligned with LLM embeddings of scene captions"

If those added instructions are live in claude.ai (as they should, given the update is dated July 31st), they do not seem to have had much effect to how Sonnet 4 answers this question. This is how it answered right now:

OpenAI: LIVE5TREAM THURSDAY 10AM PT

There was a time when changing the constitution required a vacuum cleaner ad and a time machine.

First round results: o3 beats Kimi K2 Gemini 2.5 Pro beats Opus 4 Grok 4 beats Gemini 2.5 Flash o4-mini is about to beat DeepSeek R1 This is a great way to benchmark models. Difficulty will scale beyond human levels, without limits.

Google’s Gemini app now lets anyone create a personalized, 10-page storybook with custom art and read-aloud audio—just by describing your story. Want pixel art, comics, or even crochet style? Gemini’s got you covered in 45+ languages, and you can even add your own photos for a personal touch.

I don't think they quite managed to reset the rules. One significant limitation is that these are text-only models. Which makes sense if they have made multimodal capabilities an even more significant selling point for GPT-5.

Sam Altman moments ago: "we have a lot of new stuff for you over the next few days! something big-but-small today. and then a big upgrade later this week."