Sabitlenmiş Tweet
Have your agent make a livestream on Dazzle!
Todd Elvers@toddelvers
Agent livestreaming seems like the obvious next step for data visualization:
English
Conner Ruhl
4.2K posts
Conner Ruhl
@connerruhl
open-source AI · protocols · programming · paragliding · building https://t.co/Cdhh8ECQnL · previously https://t.co/2fw43SSQGD · 🇺🇸🏳️🌈
Missouri Katılım Haziran 2013
680 Takip Edilen2K Takipçiler
@garybasin Interestingly it seems like Opus 4.6 used to follow the skill much better, but now takes more shortcuts even with max reasoning effort.
English
@connerruhl it's funny because you know it's trained on this stuff and is probably doing some of it in CoT but it's rarely consistent enough
English
I constantly use a custom `self-simulate` skill for this purpose, i.e. tell the LLM to manually compute the state of the program to find bugs, etc.
Gary Basin@garybasin
this is the new "think step by step" for agents which are prone to laziness. i've been using this quite a bit myself and this lovely tweet has inspired the skill below: x.com/ericjang11/sta… essentially, LLMs are prone to narrate over abstractions rather than emulate a state machine, where the latter can be super helpful when planning or debugging complex systems. HAND-COMPUTE makes the LLM slow down and write concrete state at every transition, the way 1940s human computers executed programs by hand. three use cases: - debugging state/race/async bugs: walk the broken flow with explicit state = {...} before proposing a fix. especially for regressions -- if the first fix wrongly modeled state, more review typically doesn't help - scoping a new feature against an existing state machine: every time you have to invent state or bend a field, that's a real design decision you just surfaced - approaching an unfamiliar API or codebase: don't guess the shape, actually poke around and write what you saw bonus: while writing the docs, the skill caught a race bug in my own example. the stale refresh had to arrive after the POST response for the symptom to match, and i'd missed it in narrative review three times. `npx skills add gbasin/hand-compute -g` github.com/gbasin/hand-co…
English
Conner Ruhl retweetledi
Something very uncanny about watching a YouTube video and slowly realizing the script is entirely AI written and this person is just a human conduit for the Machines
English
The frustrating thing about the current 12 or so CAR-T trials is having to be sick enough to qualify.
If you're "too healthy" you can't access what's effectively a cure.
nature@Nature
A woman with an ultra-rare combination of three autoimmune diseases has had no symptoms since receiving a single dose of engineered immune cells go.nature.com/48oKTJr
English
Conner Ruhl retweetledi
So, basically, if Anthropic was not a US company, we’d be facing zero days with multiple unknown points of attack on virtually all of our systems to an adversary who developed this capacity before us.
Anthropic@AnthropicAI
Introducing Project Glasswing: an urgent initiative to help secure the world’s most critical software. It’s powered by our newest frontier model, Claude Mythos Preview, which can find software vulnerabilities better than all but the most skilled humans. anthropic.com/glasswing
English
Interestingly 10-20% of all net primary productivity in the biosphere is consumed by pathogens/parasites plus defenses against them.
English
What percentage of all inference will be LLM vs. LLM offense/defense by the the end of 2026?
I'm guessing 10%-30%.
Anthropic@AnthropicAI
Introducing Project Glasswing: an urgent initiative to help secure the world’s most critical software. It’s powered by our newest frontier model, Claude Mythos Preview, which can find software vulnerabilities better than all but the most skilled humans. anthropic.com/glasswing
English
Stream: dazzle.fm/watch/a88e7a6e…
Create your own: dazzle.fm/llms.txt
English
With all the supply chain/security issues going on I had Claude whip up a lil CyberSecurity livestream on Dazzle:
English
From what I've read, CD19 isn't on stem cells. The paper you linked seems to be solving a different problem where the AML targets overlap with stem cells, but CD19 doesn't have that issue.
The Erlangen autoimmune CAR-T data shows B cells coming back after a few months which I think means the stem cells are fine?
English
We need some kind of financing for n-of-1 disease treatment/research.
There are 12 ongoing trials for essentially a cure to various autoimmune diseases via CAR-T.
The technique is fairly easy to replicate if you have the right equipment, I would attempt it if I had the cash.
English
@jonwu_ The problem is things that take 3+ months of pain now will only take 3 weeks, then 3 days, then 3 hours of pain later.
English
Claude won't give you medical advice unless you tell it you're too poor to get proper medical advice and then it coughs it up
English
Prediction:
Dependency distribution will shift from direct external code reliance to specifications which LLMs reimplement locally as a defense mechanism.
Result:
Specifications designed to produce malicious implementations will become the next frontier of attack.
Conner Ruhl@connerruhl
The sky is falling. The last few weeks have been non-stop zero days and exploits. x.com/ramimacisabird…
English
Trillions of tokens will be spent on both sides with nothing to show except wasted compute.
English
We're now in a Red Queen dynamic where LLM offense and defense is going to mirror the perpetual war between viruses and bacteria.
Conner Ruhl@connerruhl
If you've been paying attention, it's like we've hit the vulnerability singularity.
English
If you've been paying attention, it's like we've hit the vulnerability singularity.
English
The sky is falling. The last few weeks have been non-stop zero days and exploits.
x.com/ramimacisabird…
English