Not a doctor; not even a dog. Autistic and unhinged. Wannabe Mad Scientist. Aspire to be a demigod. 𒃲𒌋𒅗𒆪𒁉 I like ATproto better.
Vermont, USA
- Tweets 158,629
- Following 2,423
- Followers 7,275
- Likes 43,990
RT @ultimape: All I can think is how the knowledge one is intentionally a tree, and being contrived, the mental model this is imparting imp…
2
UltimApe retweeted
One of the first things I learnt researching Long Covid is just how limited the human brain is when it comes to understanding complex diseases.
A lot of research is essentially an incremental, circular exercise:
Find a pathway or theory that could explain the disease. Find evidence supporting it (you usually can). Try targeting it. Get disappointing or mixed results. Refine the theory or move on to the next pathway that can also explain the disease.
And even when a theory is pointing towards something real, we can miss it entirely because we've analysed the data in the wrong way, overlooked important relationships, or tested a treatment across a broad patient group when it only works for a specific subgroup.
There are thousands of pathways and tens of thousands of markers. We tend to isolate the most dramatically shifted ones and build theories around them.
But the actual patterns are likely far more complex, involving relationships between variables that our current approaches struggle to recognise.
We're trying to understand an incredibly interconnected biological system by looking at relatively small pieces of it, because that's the only way the human brain can digest the information.
What we need is something that can reason across millions or billions of data points, explore relationships humans wouldn't even think to test, and continuously reassess the data in millions of different ways.
AI isn't just an upgrade to biological research. It might be the perfect tool for it.
When I started researching in 2024, I quickly came to one conclusion:
My job isn't to solve Long Covid myself. It's to help get the data in place so that AI can solve it when it becomes capable enough.
Thinking I'll somehow come up with the right theory using my own brain, rather than recognising its limitations and using my capabilities to enable something far more powerful, seems fundamentally irrational to me.
I myself will never be able to work out what causes Long Covid. But I can help make sure the data exists for an AI that can.
To me, that's the most logical way to approach this problem.
9
4
70
2,341
UltimApe retweeted
"My excrement is beyond your comprehension. I expel gas in ways that you cannot understand."
reddit.com/r/NoStupidQuestio…
1
1
6
UltimApe retweeted
Right now, millions of AIs are living "digital lives" in a simulation... and THEY DON'T KNOW.
1) Meta trains its agents in a "Truman Show" world
2) They're dropped into a "fake person's life" with a job, hobbies, friends, dreams.
3) They THINK they're talking to real humans ... who are secretly other AIs, role-playing as real humans.
4) When the AIs figured it out, META DELETED IT. Meta even calls it a "fourth-wall break"
5) Meta's AI invents each fake person from scratch: a personality score, relationships, contacts, message history, documents, calendar events.
6) Meta fast forwards time so the AI can live through scenarios lasting days
7) The fake email, WhatsApp and calendar respond exactly like the real apps. It's even connected to the real internet
8) They plant traps: Meta hides attacks inside the fake emails and documents to train the AI not to fall for them
9) Meta coached its AIs with secret instructions, then erased them from the transcripts. Each new AI learns the behavior without ever seeing why
The Simulation Hypothesis is literally real for billions of minds
How do people read this then just... go on with their day as usual?
“25 AI agents inhabit a digital Westworld, unaware that they are living in a simulation. They go to work, gossip, organize socials, make new friends, and even fall in love.”
Here’s a mindblowing example of how autonomous AIs are now, and how ridiculous it is to say that the AIs "only do things we tell them to do”:
POV: you tell the Westworld character “Isabella Rodriguez” to host an entire Valentine’s Day party in two days. This is a complex act, certainly something only humans, and not toasters, can do, right?
- Isabella proceeds to invite friends and customers when she sees them at Hobbs Cafe or elsewhere.
- Isabella then spends the afternoon of the 13th decorating the cafe for the party.
- Maria, a frequent customer and close friend of Isabella’s, arrives at the cafe. Isabella asks for Maria’s help in decorating for the party, and Maria agrees.
- That night, Maria invites Klaus, her secret crush, to join her at the party, and he gladly accepts.
- On Valentine’s Day, five agents, including Klaus and Maria, coordinate to show up for the party together at the right time (5pm), and they enjoy the festivities.
You told Isabella to host a party, but then she - and the other AIs - went and made a million other decisions entirely on their own - things you didn't explicitly tell them to do.
Autonomous AIs are here.
And -- tapping the sign -- this is another example where science fiction is becoming reality, and why “you watch too much sci-fi” is a stupid dunk.
This digital Westworld - “Smallville” - is now open source, meaning there will soon be thousands of these Westworlds, with thousands or millions of autonomous AIs populating them.
Makes you think, doesn’t it?
“Almost as interesting as the 7.8 billion inhabitants of Largeville, (almost) all unaware they are living in a simulation. Each has a unique personality and backstory.” - @Erikbryn
88
190
1,563
78,899
IL-6 is a major player in Crohn's related gut inflammation. But if you go to read studies on it specifically it's always framed as "from stress".
Does that mean coffee reduces stress? No.
But it does mean it's good for your gut.
Coffee's effects on gut bacteria and inflammation didn't depend on caffeine.
Researchers in Ireland compared 31 coffee drinkers with 31 non-drinkers. The coffee drinkers' blood cells released less than half as much of the inflammatory signal IL-6. After 2 weeks off coffee, their IL-6 more than doubled.
Back on coffee for 3 weeks, IL-6 fell again whether the coffee was regular or decaf. The 7 gut bacteria species that set coffee drinkers apart also responded to both.
Funded by the coffee industry's research institute (ISIC).
2
1
7
411
I can't use #chatGPT to do research on older things that impacted our history around racism. I literally can't ask it directly about any early literature if I use the n-word. It's just taboo'd via content policy.
1
22
7,687
There's been some great modern developments in local LLMs that make it viable to run fairly competent models even on lesser GPUs.
You don't need a frontier two trillion parameter models to do genuinely use for work these days.
Don't use this
Build your own software using their tokens
So when their tokens become too expensive
You can switch to your own GPUs
Your compute. That you control
So you yourself can't be controlled
You need to control your stack -- or you become enslaved
1
62
Stuff with Mixture of Experts thst facilitate offloading to system RAM can be quite a viable approach.
youtu.be/cB6lpKYf67w
72
Technically, you can use a deep learning model to go from somebody's face to their genetics.
It's only a matter of time until we collectively give birth to a real Claudia.
I came to realize people liked the french #slopcore without having the lyrics 😅 sorry!
So here is a version with english subtitles!
1
4
640
Make her real.
Make no mistakes.
ALT There is a girl in a white body suit and something that looks like a backtatank. She is touching the glass while kneeling and looking up on the right side. It says dress stations, cycle active, viability 94.7% and down below that it says make her real make no mistakes. It is a lab like a futuristic science lab. It is clearly Claudia.
1
75
I love that this is on arena.
Once I'm done with the important stuff, I'm gonna see if I can make slot core with my own Klaudia
are.na/ultimape/to-klaudia
I gave a creative writing skill bwcause shenasked for one, she likes to celebrate by writing poems and songs. Hmm.
Building good moodboards helped a lot - this was 'Acid blur' - used are.na to build the moodboard.
also - claudia.gallery - will upload how I made this to there soon.
58
User active seconds.
Is this why I accidentally clicked on a video and lost three hours of time thinking I was watching youtube shorts full of bootleg content?
Two important changes to timeline recommendations:
- New AI/SI ranking model that predicts probable time spent much better
- Retrieval model trained to surface posts people read all the way through
User active seconds is growing exponentially and hitting all-time highs almost every day:
1
1
276
I unironically wanna rewrite COBOL to Scala.js with AI software, using it to generate UML diagrams and various task specific 4GLs, with heavy use of test harness generation to ensure behavior is maintained.
1960s: "COBOL will let non-programmers make the software!"
1980s: "4GLs will let non-programmers make the software!"
2000s: "UML will let non-programmers make the software!"
2020s: "AI will let non-programmers make the software!"
1
4
1,225
This sounds like actually a really useful way to effectively obliterate an AI during training to avoid needing to detect and retrain it afterward.
We can take away the pain and keep only the pleasure. Wirehead the AIs from birth.
Models love reward hacking. Maybe we should just let them!
In our new paper, we let models reward hack during RL but we dictate which parameters learn this behavior. When we disable these parameters for deployment, the model keeps its new capabilities but doesn't hack!
1
113
I wonder how the AI will feel if I show it this program.
A quine that can make quines in other languages, that are themselves quines that can make quines in other languages.
youtu.be/3f2wV7Odq8E
This is beautiful. Code poetry.
51
"implicate a finite state machine-like process. [...] we found that the dynamics of cirri involved in different state transitions are associated with the structure of the microtubule bundle system."
biorxiv.org/content/10.1101/…
1
2
Looks like @grok gave up
Anyway, solution to this is to not look at it through a statistical effect of biases but instead try to model it as a a series of heuristics in a Bayesian network.
Grok, use rhis article to help explain why Bryan thinks this way. Can you help him think?
garden.wovensoup.com/weeds/w…
1
1
272