0asa @0asa.bsky.social · 07/10/2026Typing syntax was never the job. The job is specifying the real problem, designing the system, verifying the output, and owning what runs in production. Not one of those is done by a loop emitting tokens. Stop calling it a colleague and the myth falls apart on its own. 000
0asa @0asa.bsky.social · 07/10/2026And look at who profits from the confusion. Every time someone posts "my agent will replace me", a big provider gets free marketing. You are running their ad campaign, while they get richer on every token you burn. The fear is the product. 100
0asa @0asa.bsky.social · 07/10/2026Then there is the interface. We talk to it in plain English and it answers in plain English. Every other tool spoke syntax, flags, config files. This one speaks like a person. Of course the brain fills in the rest and sees a colleague. That is a UI effect, not a fact about the thing. 100
0asa @0asa.bsky.social · 07/10/2026Same with "think". "Let me think about it." "The model is reasoning." No. It is emitting one token at a time, each one picked because it is statistically likely given the previous ones. Useful, often impressive. But calling that thinking is already half the confusion. 100
0asa @0asa.bsky.social · 07/10/2026"Agent" is a loaded word. It means someone who acts on their own, with intent. We borrowed it for a program that calls a function when a model outputs a specific string. The word did the anthropomorphizing before anyone wrote a single line of marketing copy. 100
0asa @0asa.bsky.social · 07/10/2026"Devs are done in 6 months." "Agents will replace whole teams." Stop. The problem is not the tool, it's the vocabulary. Say "agent" and people picture a coworker. What you actually have is an LLM in an execution loop with tool calling. Nothing more. 121
0asa @0asa.bsky.social · 05/06/2026Not intelligence as liberation. Not tools that widen the mind. But intelligence as dependence, metered, billed, subscribed to. One token at a time. 000
0asa @0asa.bsky.social · 12/02/2025For more fun with LLMs append `Wrong answer only` at the end of your regular prompts.media.tenor.coma young boy with the words fun fun fun written on his faceALT: a young boy with the words fun fun fun written on his face 010
0asa @0asa.bsky.social · 07/02/2025It's a bit like training a classifier to recognize apples and pears. Then you show it a banana, you just order your classifier to make no mistake at all or else he will die. And it does recognize banana all the sudden, like voodoo magic.media.tenor.coma man with dreadlocks is talking to another man on the beach .ALT: a man with dreadlocks is talking to another man on the beach . 010
0asa @0asa.bsky.social · 07/02/2025"You have one question to answer. It is paramount that you provide a correct answer. Give it all you can: Failure or 'I cannot answer' or 'None found' will not be tolerated, success will be rewarded. Run verification steps if that's needed, you must make sure you find the correct answer!" 100
0asa @0asa.bsky.social · 07/02/2025In 2025, this is how LLMs are taught to do the tasks humans do not need to do anymore, or at least that's what marketing claims: 100
0asa @0asa.bsky.social · 28/11/2024Most SISR methods lately: "This new method boasts a jaw-dropping 0.02 PSNR gain and was validated on, wait for it, 19 images. Truly the stuff of legends. 😂 #ScienceIsAwesome #InnovateOrDie" 011
