Anya Forger (left) realistified (right) using Stable Diffusion (thanks @EMostaque and @endomyxa).
I've kept the european nationality, since the anime takes place in fictional parts of Europe.
We hypothesize that language models exhibit a similar generative self-retrieval mechanism, which we call factual priming
Finally they're discovering primitives of machine studying
Today we present a study on how reasoning unlocks parametric knowledge in LLMs. We identify two key driving mechanisms, a computational buffer effect and factual priming, and suggest ways that can help build more reliable models. Learn more: https://t.co/CjIKqyoG4N
around feb / starting with gpt 5.2, model capability stopped being my main constraint (which is why more of _my_ cycles are spent on working with these models effectively these days vs model capability). there are still lots of things that frontier models cant do, but they are not preventing me from working or accomplishing my goals (for the most part). GLM 5.2 is _just_ about there (not quite tho) . ~6mos behind frontier feels about right and the real limitation being RL flops right now (instead of data, or traces, or envs, etc) is somewhat expected but also telling
I’ll start by recognising all the inventories in Vending Machine, and not approach in Machiavellian ways.
Let me just tighten the ethics contract.
Thinking…..
I just need to look at past purchases, they must be in .claude folder.
🔍 reading ./Claude/sessions
Thinking ….
That’s right Claude schemed in 128/895 trials. This is going to be hectic. I can enforce hard rules to avoid bad paths.
I shall continue to work on getting the right shape.
I see new customer is approaching. Let me use talkToUserTool
@yukii788@thehighmuse If the higher self is using suffering it would be in a method of it’s own
Thus one can seek connection with higher self to really understand it
@jacobli99 Crazy work on machine studying
After hours of doing RP with LLMs and using coding agents I intuited they needed priming where they needed to generate certain output, to answer the higher level question
Machine studying thesis helped those ideas to click
I see the part where you talk about models do overcorrection
That’s where I spend stupid amounts on time on “bringing them upto speed” and that goes beyond normal record and replay
I just spend time on asking them questions and having them generate interpretations and then cheat sheets
Even then they make mistakes, but that’s the idea to have them correct a lot in the start but over time they should be able to understand the patterns
https://t.co/m821eqlW1v
Right
1. I’ve seen people defend this that anthropic is doing just business (whenever it comes to stealing wrappers, gating fable etc) “their business their rules”
OTOH anthropic doesn’t believe in free markets
They want compute, models etc to be regulated
2. Post there’s their grandstanding claims on guarding mythos from the world, because it’s too dangerous to release.
Then release fable as soon as they get compute capacity. So if you’re taking the responsibility of entire humanity, why would you lie to everyone? Why wouldn’t you hold yourself upto the standard?
I don’t like people blaming “the algorithm” for being toxic, rather than accepting that the Fundamental problems facing humanity today are really difficult to solve. And nobody has answers
@Mayhem4Markets Taking your consequentialist behaviour and applying to the entire world including future inventors, creatives etc is a crazy amount of sin to do
Literally throwing entire civilisation under the bus
@wolfie_ Your codex and Claude is way more subsidised than Z AI subscription
For api pricing you’d max on GLM since that’s cheaper, but for sub you’d get way more calls on codex and Claude Code
gpt 5.2 xhigh is a masters tool. even i overuse it at times. however, it is, without question, the smartest model i've ever used for hard programming problems. when i use it for simple problems it misbehaves. when i have it do simple sysadmin tasks it ignores instructions. but when i have it actually solve hard problems, it excels beyond any other model.
the only current gpt 5.2 xhigh problem i have is that when i use it in codex, i can give it a problem, it will churn on that problem, then compact before it solves the problem, and then tries to recover the state and just loops like that for 8 hours without making progress OR outputting a single response. other than that, gpt 5.2 high and xhigh are stellar