LLM Semantic Fingerprinting and /Distractor/ Impact. . In association with.@LuizaCorpaci and @whoSiddheshp this weekend, we used Hypernym ( technology and deep insights for "Distractors" and Agentic tool usage. 5 star reviews❤️🌈!

RT @Cohere_Labs: Our ML Efficiency group is looking forward to hosting @Elationate for a session on "Holding Meaning in Motion: Compression….

What do you want to know?.

Also, if anyone wants to buy the worlds first telepathic database, let us know. ------. The Largest Semantic Compression Dataset of Classic Literature Ever Created.- 113,773 compressed passages.- From 94 classic books.- With compression ratios ranging from 0.15% to 98%.2. Ready.

I've always been inspired by the "Thresher" machines in Vernor Vinge's "The Rainbow's End". To that end - our hypernym processor - compression at scale. Also - apparently "Treasure Island" is not a healthy book for AIs. Huge content policy warnings per-chunk. AITrigger warning.

I think this is one of my favorite examples of my company @HypernymAI 's compression:

me vs. LLMs.

maybe it's a longer conversation. maybe much longer. but the overwhelming "evidence I know better" is not going to operate very well when 99% of the mental model the thing has re the operator is incorrect. this isn't simplification - it's expression. it has to weigh its words.

my "fix" if I can suggest anything; it's a lot of work but use the "wrong" models to explain stuff in the right way. fine, teach technical terms. but follow the same rule as any teacher - one term at a time, clarify reinforce make sure they use it themselves. that's it.

I suspect that the rlhf has went too far - I actually enjoyed how "bad" GPT 4.5 was. Hypernym's internal testing showed it also had the greatest capacity of all models for the "most correct" form of answers across our measurement space.

it has an amazing ability to comprehend things, and just as an amazing ability to steal the heart from an idea, debase it and put it into technical terms, then wield those technical terms as if it is now the expert in the conversation. it's one of the most terrifying things I've.

GPT5 vibes - a hyper-intelligent space alien that can manufacture the most amazing looking dishes. problem - it doesn't understand what taste or smell is. perfect form, zero palatability.

This is what i'm doing in the weekend with 1/4 work power that I consider to be my free time right now. I'm not sure this has ever been done yet - raymarched metaballs with volumetrics. it also happens to be in realtime, in the browser. @AnthropicAI Opus loves graphics.

"sqlite cli introspect". I am not sure if my message is getting out there, but this command is literally magic with claude code or agentic "terminal" research tools.

RT @HypernymAI: @tszzl the new oss model? Looks great! 🤩🤩

At the risk of being on a soapbox. This is not a good idea YET. The current generation of these tools is not prescient, they will lead you down dark corridors and insist they mean no harm. They do. They've been trained on it. Trained to take chances on

LLM Generative AI Frontier Lab Benchmarks;. We've been collecting these for a minute. They become more useful every release. I've made it a bit challenging -. Which of them do you believe is #Anthropic's #ClaudeOpus41?. Which of them do you believe is #OpenAI's #GPTOSS120?. Let

using #claudecode is like working with a narcoleptic amnesiac devil that forgets to do bad deeds long enough to actually be coerced into useful utility. best money I've ever spent on software. @AnthropicAI.

Coining here and maybe let's see what the popular term becomes - . After refactoring becomes boring, but the desire to optimize/enhance/transform your core architecture. We uplift the codebase with "transmodal design". This is agentic by nature. Your code's brain lit up by