wongarsu

23 ms·

451.
by wongarsu 5mo ago
This is their only New Glenn launch pad, but the pad for the 9x4 is already under construction. Depending on the damage sustained to the pad that might be a factor in the decision
452.
by wongarsu 5mo ago
I believe that's the inevitable outcome. The Chinese will build a moon base, as a sign from the Chinese government to the Chinese people that China is capable of cutting-edge engineering and science (notably a demonstration to their ow
453.
by wongarsu 5mo ago
False is correct, but misleading My implicit assumption is that if you fact-check the fact-check, any label other than "true" means the original fact-check is unacceptable
454.
by wongarsu 5mo ago
A major limitation is that they only test GPT 4o. Previous research like [1] investigating the same question has shown significant differences between models, and even depending on the language of your prompt 1: https://aclanthol
455.
by wongarsu 5mo ago
Of the available options, "Misleading" is probably the best, since something that is most likely true but unproven is presented as fact But "unknown or undecidable" should have been a category.
456.
by wongarsu 5mo ago
The video thumbnail implies bouncing off the ground, not a wall. Not sure how the geometry works out for that
457.
by wongarsu 5mo ago
One fun example: "Ruskin Bond was born on May 19, 1934, in Kasauli, Himachal Pradesh, India". Opus and Gemini believe this to be true, GPT 5.4 believes it's false, Sonar thinks it's mostly true. Disagreement value of 3,
458.
by wongarsu 5mo ago
Yes, the labels are weird. Most misleading statements are true. Any "mostly true" statement is false. I suspect the intention was "Factually true, and no gotchas exist", "technically not true, but so close to the tr
459.
by wongarsu 5mo ago
According to the benchmark it is. "Only one verdict bucket can be correct per claim, so any disagreement among the panel means at least one model's verdict is label-inconsistent under this 4-bucket rubric (True / Mostly True
460.
by wongarsu 5mo ago
Like this: https://c8.alamy.com/compde/t0580m/der-verkehr-kurve-spiegel... Or this: https://cdn-01-artemis.media-brady.com/Assets/ImageRoot/DMEU... Reasonably common in difficult corners
461.
by wongarsu 5mo ago
AMD also had the strongest offering for GPU and CPU using the same memory with the same address space. That allows you to switch between CPU and GPU processing for the same data, without paying the cost of moving the data to and from the GP
462.
by wongarsu 5mo ago
But the feature list at https://claude.com/pricing#team-&-enterprise literally lists "tiered incentives on committed spend" and "non-standard terms" as perks of the sales-assisted Enterprise plan. Ma
463.
by wongarsu 5mo ago
Back then people were mostly farmers, but we already automated that job away. Not completely, but compared to the middle ages we 50x'd their output. Which is a great illustration what it means to make a job 50 times more productive. We
464.
by wongarsu 5mo ago
If you point an LLM at a middleware and ask it to find vulnerabilities, then not finding this is a shortcoming. Whether "LLM failed to spot vulnerability that took humans 8 years to find" is a great headline about shortcomings of
465.
by wongarsu 5mo ago
The posted page has an entire section titled "Why didn't Mythos find this?" tl;dr: the bug spans three components in different code bases that when looked at in isolation each do reasonable things. The bug is in the interacti
466.
by wongarsu 5mo ago
Other great additions: - Left and Right parenthesis with middle ring [1] - A wiggly exclamation mark expressing mirth or laughter [1] (edit: and something I completely missed: the inverted version can express sarcasm) - Cuneiform numerals,
467.
by wongarsu 5mo ago
> if not is there really a point to include everything possible in unicode ? Needing to load three fonts to show a single document that mixes vastly different character sets is still infinitely better than not being able to have those di
468.
by wongarsu 5mo ago
There we go back to the original question: are subscriptions profitable, API pricing wildly profitable, and they just lose all that money on fixed costs like model training; or do they actually barely make money on inference? That's wh
469.
by wongarsu 5mo ago
Sell more. The hope is that there is a huge addressable market that includes huge per-worker demand in almost all white collar work and lots of inference in people's private lives If that doesn't work, then yes, then prices will h
470.
by wongarsu 5mo ago
If replacing github wholesale isn't viable, how does the story for replacing GitHub Actions look like currently? I don't remember the pre-Github-Actions days of everyone using CircleCI with a github integration in a negative light
471.
by wongarsu 5mo ago
Selling inference for more than inference costs is not incompatible with bleeding cash at ungodly rates. They do in fact pay ungodly amounts of cash for other things, like training, marketing, etc. Heck, you can bleed cash while being profi
472.
by wongarsu 5mo ago
I am conveniently not counting training costs (since they add no marginal costs, selling more tokens doesn't impact them), and hardware and DC costs only amortized Of course they do have to "make bank" in some way to offset t
473.
by wongarsu 5mo ago
Depends on what their actual costs are. Either they are losing lots of money on subscriptions, or they make absolute bank on API pricing. Looking at the pricing of 1-2T models like Kimi or DeepSeek on the open market, I'm tempted to as
474.
by wongarsu 5mo ago
Absolutely. We are in a phase of "free money" for AI. Just as with the dotcom bubble that leads to 1) lots of experimentation, and 2) lots of infrastructure buildout (which includes AI model training). Once the money dries up, som
475.
by wongarsu 5mo ago
In 2013, the same deal would likely have gone through. US-Dutch relations looked very different in 2013 under Obama than they look now under second-term Trump. Any reciprocity today based on things Obama did back then falls flat because we
476.
by wongarsu 5mo ago
I'd assume mostly by avoiding the need for frequent rebuilds. Incremental builds are pretty fast (at least fast enough for my needs on a moderate codebase), full rebuilds can be brutal There are also some optimization tricks related to
477.
by wongarsu 5mo ago
And to someone trying to make better lawn mowers it would be very intetesting to study that pattern and deduce its causes. Same for the golf course owner who wants better results from their lawn mower, or the enthusiast trying to understand
478.
by wongarsu 5mo ago
That's what you'd expect. But we don't know for sure why GPT4.1 chooses 69 only a quarter as often as a random dice roll would. And we don't know if this quirk is reverted by 'uncensoring' a trained model
479.
by wongarsu 5mo ago
If you end a search query with a question mark, kagi answers with their version of search overview. But with a quality closer to asking an agent with access to your search results. It's great for one-off queries
480.
by wongarsu 5mo ago
Overengineering and convoluted code stand out when reading. The hard part are the subtle errors. And the Rust compiler helps you out a lot more here

More ›