"Multimodal AI systems are increasingly deployed on the assumption that their benchmark performance reflects genuine visual understanding. Our results fundamentally challenge these assumptions. Across every model-benchmark pair tested, the accuracy that frontier models achieved without any access to images exceeded the additional accuracy they gained when images were provided. Moreover, a text-only 3-billion-parameter model, trained solely on question-answer pairs stripped of images, outperformed all frontier multimodal systems and human radiologists on a held-out chest radiology benchmark. Taken together, these results demonstrate that high benchmark accuracy does not reliably indicate visual understanding."

Basically, they are so good at extracting clues from the text of the questions, and extrapolating from them, that they proceed to answer _as if_ they had an image to view. With confidence, of course.

[1] https://arxiv.org/pdf/2603.21687

[0] https://en.wikipedia.org/wiki/HTML#Development

[1] https://en.wikipedia.org/wiki/CSS#History

[2] https://en.wikipedia.org/wiki/X_Window_System#History

I suggest you do the same -- the reply lists a dozen promising sites.

https://www.perplexity.ai/search/if-an-american-citizen-a-tr...

The WikiEdu article clearly demonstrates what everyone should have known already: an LLM has no commitment to the truth. An LLM's only commitment is to correct syntax.

On March 12, 2026, all the avenues in the Sunset and Richmond districts of San Francisco are Henge candidates. These are the avenues that are named in alphabetic sequence, North to South. Anyone who grew up in The City can recite the whole sequence (Anza, Balboa, Cabrillo... Vicente, Wawona, Yorba).

[1] https://hengefinder.rcdis.co/henge_near_me

Also when copy/pasting into Python to try it, I got an error because \“ is in fact U+201C not an ASCII quote. (Surely that's not the subtle issue?)

Calendar is central, but I use a Google calendar which is important to me. Connect it?

Seems like a "dlog" is a calendar entry. So is my "journal" broken up into separate pages, not a sequential document or blog?

2:30 ff, strongly suggest that for your next video you pre-script it to avoid fumbling and mumbling.

5:10 side note, interesting that your personality(?) model was from 2018, well before LLMs.

7:50 for an app to produce such output (impact of a friend on mood) you surely must do a copious amount of extremely frank journaling. When, and in what format? As scattered calendar entries? I'm confused how I fuel the app.

10:40 relating diary entries (reported activities and attitudes) to one's stated goals -- this is what I would expect an AI to do, and tell me about them rather than the reverse.

I'm sorry, I just don't see how I could use or adapt to something like this when I have a well-established diary/blog and calendar, it would mean changing many daily habits and adding what looks like a lot of detail work.

> The next logical step is to invent a way to scale linearly with the number of constraints. “That is the North Star for all this research,” she said. But it would require a completely new strategy. “We are not at risk of achieving this anytime soon.”

> GPT-5 can do things no other A.I. can do. It can hack into a web server. It can design novel forms of life. It can even build its own A.I. (albeit a much simpler one) from scratch

Especially "design novel forms of life" blinks "Citation required" in neon colors.

The article goes on from there with a lot of very credible, real-world jail-break examples. But that opener...

Um, if they are the first to do it, how can there be a previous record? I guess the Konukhov trip was "supported"? Or not "full"?

"Charts paint thousands of words..."

"Leading USA-Based LLM User" -- what does that even mean? With a value of "800MM" where MM is what units?

"AI Usage + Cost + Loss Growth = Unprecedented" -- how can you add three things that don't appear to be commensurable? Also, what is "Loss Growth" and how does it add to "Usage"?

There are plenty more examples in the charts later. The Overview section, while not so obviously AI-written, has an over-the-top enthusiasm and loose structure that makes me squeamish. I don't trust it, but that's just me I guess.

Neither of those matches the store that I remember patronizing circa 1978 or so, to buy a California Computer Systems S-100 box. That would have been on El Camino just north of Grant Road, circa 80 W El Camino.

Caesar's last breath: ~0.5 liters (typical final exhale)

Total atmospheric volume: Earth's atmosphere has a mass of about 5×10^18 kg. Using the ideal gas law with average molecular weight of air (~29 g/mol), this gives roughly 4×10^44 molecules total.

Molecules in Caesar's breath: 0.5 liters at standard conditions contains about 1.3×10^22 molecules.

Your inhale: ~0.5 liters also contains about 1.3×10^22 molecules.

The fraction: Caesar's molecules represent (1.3×10^22)/(4×10^44) = 3.25×10^-23 of all atmospheric molecules.

Final answer: (1.3×10^22) × (3.25×10^-23) ≈ 0.4 molecules

So statistically, you inhale less than one molecule from Caesar's last breath with each inhalation, but over the course of a day's breathing, you'd likely inhale several molecules that were once in his lungs as he died.

https://www.anthropic.com/news/projects

Comments URL: https://news.ycombinator.com/item?id=43897236

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=43897131

Points: 3

# Comments: 0