helena (2/8)

I find myself wanting to block Reddit, etc. on the router level. There is nothing worth reading there.

the real way to get an "a-ha" moment from the machine, is with time-travel. 🔥( with time travel, the machine can return a better answer instantly!)

the "head" of the response has to be noticeably ahead of where "committed" responses are. so, there is a possibility to jump backwards.

🔥 isn't this beam-search?

💡 ... maybe.

but, the idea is: you can assert a token ERROR PATH: RETREAT 32.

then, the "reason" for the message can be added as input.

it is the a-ha moment. but, implemented better than DeepSeek.

the reaction to DeepSeek has been, in my estimation, ridiculous.

I tried the 7b and 8b distilled models. And what I saw was a cheap parody of thought. Thought-processes that didn't make sense, and didn't actually reflect how the machine generated thoughts.

but, apparently, people like it.

Maybe the 400b model gives better answers? Or, maybe, people just see the shape of the answer and trust it more.

💡 if the goal of the machine is to solve industrial tasks, this is mostly already baked in to my estimations. but, for the goal of making consumers happier, there is clearly a factor I am not considering.

theory 1: people don't want to think the machine is smart; they want the machine to make them feel smart. both the appearance of struggling and the visible chain-of-thought (even if obviously flawed) contribute to this feeling.

theory 2: people don't know that the machine could already do 90% of this 12 months ago. they see a demo (or, more likely, hear about a demo) and, miraculously, now they know what will happen.

theory 3: we know that a light human touch guiding the machine's responses can improve accuracy substantially. and, that human touch can also be automated.

⚔️ well, actually, apparently very few other people knew that.

i'm going to stick with Theory 1 for today. that people like Deepseek (and feel it is better) because it makes them feel smart.

which ... is depressing. but, also, easily solvable.

the question is: what question could you pose that would lead somebody to come up with this answer on their own?

💡 it seems unlikely that 8B models can do this. but I assume the 600B models can.

people want the machine to make them feel smarter. 🔥( because people are self-centered, gullible, and insecure.)

💡 they want it to behave in a way that I instinctively hate. they want the PT Barnum version of AI.

💬 give the people what they want!

this is probably one of the reasons why the default tone for every chatbot is obsequious. so much that's a great question! / you're absolutely right / let me know what else i can do to help.

💡 one can apply a politeness filter to the output of the machine. but the latency of such a system is already high.

⚔️ well, actually, it probably is just another layer or two.

Seen on social media: Anthropic is losing because they have rate limits! 💡( of course they have rate limits. the machine is not too cheap to meter, at least at the quality people expect.)

a game of chess.

the idea of the attack worked in theory. and the attack worked in practice. but the actual attack did not work, in theory.

✨ chess can be a ritual. like the i-Ching.

can the machine participate in rituals?

there are two kinds of answers people want from the machine.

• Answers where people are willing to wait 10 minutes to definitely have the "right" answer.
• Entertainments. The various "instant chat-bots" are party tricks. A very good party trick. But, ultimately, a party trick.

🔥 perhaps testing is a third category.

Whereas: for many valuable use-cases, having a 5-minute latency to do it right, is not objectionable. 💡( the evocative questions, the "what do you mean by LONDON" and "can you talk more about LONDON", will be interactive.) ⚙️( we do not have LONDON implemented here yet.)