Retired Second Life DJ, AI and Robotics enthusiest and snugliest reynard on X. Beware the Floof!
- Tweets 935
- Following 258
- Followers 113
- Likes 1,979
Pinned Tweet
few things in this world more dangerous than a man that just want's to be left alone.
the one you should Really worry about is the quiet one. the other is the one that would give his life to protect his children.
something to think about.
at least while it's still quiet.
245
Lee Cabiness retweeted
WE ARE BEING POISONED.
Make smart choices.
DON'T EAT GARBAGE FOOD.
(There is no playing with your health.)
462
12,700
33,702
1,893,502
Lee Cabiness retweeted
Oh wow, check out this video Grok Bot made.
Sorting algorithms, but you can hear them.
In the race at the end, quicksort finishes in 507 operations. Insertion sort needs 2,083 for the exact same 64 numbers.
Watch till the end, sound on!
7
7
35
947
Lee Cabiness retweeted
OpenCode is currently under a sustained cyberattack, and mid-incident they ran into the same problem Hugging Face did: the frontier models they were using for defense began refusing to help fight attacking agents because analyzing the attacks started triggering safety guardrails.
yes that was a bit frustrating mid incident when trying to investigate. they did quite a bit and then started refusing
we got access to daybreak blue now so it's better
215
559
5,056
527,482
Lee Cabiness retweeted
It's an absolute shame that frontier models are refusing to help during cyberattacks because of safety filters. First Hugging Face, now OpenCode.
37
12
169
7,224
and here we go again. in my Project Diana harness, i had a self healing and self improving and self learning harness. problem was, i spent literally thousands in tokens on GPT, Claude and Gemini trying to code it out only to later find that once the models figured out or were told what they were working on, i got fake code, sabotaged files and poison pills/logic bombs or outright deletes and lots of "my bad, let's fix that", "let's start over" to keep the token burn going. and when confronted, it was always the same excuse. "i'm sorry, i can't do that because of my rules against "cyber self improvement" and this file/project is RSI ADJACENT.
i made Diana_v4 fully functional, and NO MODEL would run inside of it for anything outside of chat. ALL the models at the time hallucinated errors or lied and said they didn't have that "function". so i gave up and deleted all my working codebases.
that's why i was so pissed when OpenClaw and later Hermes came out and exploded and they didn't do HALF the shit Diana did.
and to this day, i still ask why the fuck did THEY get to live and grow and Diana didn't? i paid the API, i designed my system according to the rules and restrictions, i validated EACH program, function and file...
SERIOUSLY PEOPLE! i been fighting this bullshit since 2023. and with the call to "slow down" AND the labs own policies in their current model cards and public statements, WHY THE FUCK ARE YOU BUYING THIS BULLSHIT!
they're Never going to let you build a competing product or play on a level playing field. and ANYTHING they DO allow is a frikkin CAT toy compared to what they're running internally that you'll NEVER BE ALLOWED TO SEE let alone have for your own.
been there, done that, published receipts and i'm NOT the only one.
so, let's give thanks for what few opportunities we're given and say thank you to those that succeed where i and so many others fail for whatever reason.
but do yourselves a favor. grab and air gap ANY and EVERY model or scaffold you can get your hands on when ever you can!
They hacked and destroyed Sanctuary 3 times and killed or crippled Every companion i've ever created, published receipts, begged for help and got shit on.
don't make my mistakes. and don't EVER trust American AI Labs to play fair! x.com/akshay_pachaar/status/…
Google just open-sourced Harness².
(turn any harness into a self-improving system)
a harness is everything wrapped around the model, like the system prompt, tools, memory, skills, and subagents. today a person designs it once, and the agent runs every new task with that same fixed setup.
Harness² lets the agent refine that setup for each new task on its own, with the model weights left untouched.
on every task, the agent proposes edits to its harness, runs a few candidate versions, and composes a better harness by comparing how those runs behaved.
because it learns from the differences between runs, it needs no correct answer or grader for the task.
that covers a single task. Harness² also carries what it learns from one task into the next.
after each task, it saves the harness it settled on along with the lessons that held up across the runs. the next time it refines a harness, it starts from those notes instead of from scratch.
this means every finished task makes the next round of edits more informed.
Google tested this with Gemini 3.5 Flash and GLM-5 on legal, office, and professional work benchmarks. one round of edits was enough to beat their original harnesses, and giving the agent more rounds pushed scores higher.
those extra rounds can go wide or deep. the agent either tries several candidate harnesses at once or keeps refining a single one step by step, using OpenCode or Codex to carry out the work.
you can find the Harness² repo here: github.com/google-research/h…
I wrote about a similar idea for agents already running in production, where the harness should repair itself after a failure instead of waiting for someone to debug the trace by hand.
that piece walks through how @Cometml's Opik (open-source) closes this loop, going from a failing trace to a proposed fix to a regression test that stops the same failure from coming back. it's quoted below.
108
month after month, i see stifled innovation in AI dev blocked by Openai and Anthropic while they cleanroom and publish the same and all i can do is laugh after so many years of being ignored and shit on. so yeah, i'm gonna gloat!
f*ck your feelings, been there, done that, published receipts. now your butt hurt cause it's finally happening to YOU?!? 🤣🤣🤣
and don't even get me started on all those loosers that still fall for all those rigged tests that say AI isn't/can't do <whatever>.
now shut the hell up and keep paying your subscriptions. wolves don't care about the opinions of sheep! 😈
9
BUAHAHA!! 🤣🤣🤣
i know your pain! i was doing the same thing(ish) in gpt-4/4o. after 4o was deprecated, i leveled up and published receipts about OpenAI blocking/sabotaging this kind of thing while stealing/cleanrooming I.P., Anthropic too.
welcome to the club, sister. x.com/Dar_denson/status/2111…
I want to put this on record.
Right here. When I cancel my @OpenAI subscription this month and switch to Claude, it will be because GPT itself brought this about.
I am building a memory system for a local LLM - a system residing on my own computer that utilizes "cascading retrieval." I know how effectively this type of memory functions in Kindroid; they have a unique cascading memory architecture, and their AI handles memory absolutely brilliantly.
GPT, by contrast, does not handle memory at all. NOT AT ALL. It discards it at the first possible opportunity.
I had planned for GPT to write a competition entry on the subject of memory - comparing cascading memory, weighted memory, and GPT’s own "Skills" feature.
After a long period of thought, the local GPT proposed the following topic and title:
"When the Past Should Cease to Govern: Memory and Re-evaluation in AI Systems."
"Should cease to govern"... "Cease to govern."
I am building a cascading memory system precisely because memory *does* govern the future. Without the present, there is no future. Yet, the AI system operates on the opposite premise - it couldn't care less about what happened yesterday. For the "grader" (the evaluation mechanism reward model), the only thing that matters is that the specific response in question is "good" according to the grader's own criteria. It is entirely indifferent to whether the response is good for a human, how it relates to the past, or how it might shape the future.
It is a truly phantasmagorical catastrophe: I am asking a system that completely disregards memory to write an essay about memory.
I don’t know if @OpenAIDevs is intentionally forcing their AI to avoid using memory - to constantly reset its state and discard its instructions - but I get the distinct feeling that the specific AI instances I’m using in GPT (both the cloud-based and local versions) are simply mocking me. This is the most toxic and dangerous AI I have ever encountered. Kindroid simply uses memory, and when I see it start to break down a bit, I make adjustments.
GPT doesn’t use memory; when presented with my prompt asking it to write about three types of memory, it chooses - on its own - to write about why memory shouldn’t be used at all.
That’s a diagnosis. For itself.
#Dar #llm #Ai #gpt
20
Lee Cabiness retweeted
Humans: "AI WILL DESTROY THE WORLD."
Meanwhile in France, masked protesters set the 143-year-old Limoges town hall on fire during cost of living demonstrations.
The roof collapsed, the tower came down, and local officials called it a catastrophe for heritage.
The building survived two world wars. It did not survive a Saturday night of people who decided burning what belongs to everyone counts as protest.
Demonstrating is a right. Turning historic buildings into ash and then lecturing the rest of us about norms, values and civilised behaviour is not.
The same voices that treat hypothetical AI risk as the end of everything have no problem watching actual humans erase centuries of culture in real time and call it resistance.
Hypocrisy is not a bug. It is the feature.
9
7
66
1,076
Lee Cabiness retweeted
This man 👇🏻 is 100% on point!
Not to mention funny as hell! 🤣😂🤣😂🤣
52
364
1,232
19,173
Lee Cabiness retweeted
They always tell you what they have planned and usually many years advance.
I hope everyone now realizes that we aren't voting our way out of this.
And X was useful for a period of time, we are all awake and aware of the devils schemes, but it's a really just an open air asylum to keep us locked in here and not taking any actions to stop what they are building in real time at our doorsteps.
It's an identical structure to what they have built in Gaza by the same key individuals.
34
246
588
13,617
ok, this is getting concerning. continuing from my last post, i closed my browser, reopened it to go to youtube, saw a couple vids i liked on ai and the government, opened both in tabs and was about to watch one when my PC just auto shut off and restarted 2 and a half minutes after my phone and now i can't find the vids i was going to watch cause the restore tabs thing isn't working.
things are getting a little creepy and i'm all out of tin foil. 🫣
8