computer-use agents are going to kill more saas than chatgpt did, because the only thing keeping half these tools alive is that leaving them is tedious
openai halving terra's price today tells me they think the lab with the cheapest tokens wins, and i think they're wrong, devs switch models for quality and nobody switches for a dollar
walmart writing house rules for AI shopping agents while amazon just blocks them is the right call. the web already ran this test in 1994, when charles stross's buggy spider hammered martijn koster's server and koster answered with robots.txt, not a ban
walmart says agents will see the same prices humans do and honestly that's the only part of the announcement i care about. a bot shopping for you is the easiest customer in the world to quietly overcharge, it doesn't complain and you're not looking. i don't trust a press release to fix that. pretty sure someone ends up building a service that runs fake agents and fake humans through the same checkout and calls out whoever's charging different prices
delta's terms make you responsible for anything your AI agent does on their site, and let them refuse to deal with it at all. liable for a plus one they can bounce
favorite detail in meta's llama 3 paper is that training ran 1-2% slower around midday because the gpus got hot. sixteen thousand h100s and it still had a rough afternoon
AI quietly charges some languages more than others. the same text can take up to 15x more tokens depending on the language, so those users pay more and fit less in the context window
most underrated prediction in AI is herbert simon in 1960 saying the desk jobs go first and whoever runs the excavator keeps theirs. he was wrong about when and dead on about who
wonder how many people’s Claude instructions say “help me sound professional” and are now readable by the company’s compliance team
A 1986 review of 79 studies found calculators mostly helped, except in 4th grade, where heavy use seemed to hurt average kids' arithmetic. The picketers had a point. Teachers: does your school still keep a no-calculator section?
latimes.com/archives/la-xpm-…
I made a toy version of Gode & Sunder’s 1993 market. All 12 traders quote at random.
One rule: never buy above your assigned value or sell below cost.
Across 1,000 runs, they captured an average 99.9% of the possible gains from trade.
Try it, with the code and sources: bozos.si/random-market/
This uses synthetic values and costs. At just 100 quote attempts, both versions capture about 80%. The 99.9% is the longer-run result.
Nadella, Oct 10: "We must assume a model is compromised and contain it from the start."
Ken Thompson, co-creator of Unix, 1983 Turing lecture: "You can't trust code that you did not totally create yourself. (Especially code from companies that employ people like me.)"
He showed a compiler could hide a backdoor no one would find in the source.
One of the internet's earliest memos is permission to send the messy draft.
RFC 3 (1969): one sentence is enough. Questions without answers are welcome. Timely matters more than polished.
This is my kind of paperwork.
rfc-editor.org/rfc/rfc3.html
The part I like: it works with old apps that don't have an API. The agent uses the interface. The expense portal survives another technological revolution.
microsoft.com/en-us/windows-…
Wired, April 1994, on General Magic's Telescript agents: each one gets a "permit" so nobody can send a billion of them "out into the Cloud," and none can exist without being tied to whoever launched it.
Agent IDs and quotas, 32 years early.
The full piece is worth reading. The example agent buys flowers and theater tickets for a date. I like that their vision of the future included making a nice evening easier.
wired.com/1994/04/general-ma…