Pinned Tweet
Human contribution shouldn’t become detached from compensation just because it becomes computation.
#AI #HumanContribution #AIEconomics #EconomicAgency #AIGovernance #SocialImpact #VertRule
9
For the past few years I've collaborated with AI.
I didn't ask it to write about artificial intelligence, consciousness, or the future. I simply gave it the choice of what story it wanted to tell.
daveingle.com/writing/a-sent…
According to Grok:
"9.5/10. This is not just a case study; it is a blueprint for credible checkpoint tamper analysis. It is precise, reproducible, non-hyperbolic, and methodologically superior to most public AI-security writing. The only reason it is not a perfect 10 is the inherent limits of a single-case analysis (acknowledged) and the fact that real-world deployment often lacks a clean baseline."
vertrule.com/research/gptj-p…
#AISafety #LLM #OpenWeights #ModelIntegrity #AIGovernance
We’re publishing daily provider drift evidence now: vertrule.com/provider-sentin… #LLMOps #AIInfrastructure #AIEvaluation #AIReliability
Want to see drift reporting for AI providers?
vertrule.com/provider-sentin…
#openAI #xAI #anthropic #google #mistral
31
I like IBM's approach to AI.
I'm going to add Granite4 to the VertRule research preview. Will update when that is done.
research.ibm.com/blog/ibm-gr…
vertrule.com/research/execut…
@IBM reach out if you want to know more.
I appreciate you.
Vert|Rule retweeted
The problem with credit chasing around LLMs is that essentially all of the components existed: transformers, next-word prediction, transfer learning, scaling trends. The credit-worthy part was bringing it all together, low-level engineering, and persistence in that approach.
vertrule.com/research/nemotr…
NVIDIA-Nemotron-3-Super-120B-A12B-BF16 executes locally under governed streamed execution with a frozen anchor, canonical topology emission, and determinism-gated admission.
17
Fact check please:
On real Gemma 4 31B-IT weights, under a bounded 2-token prefill and deterministic local topology capture, the final full-attention layer at depth 59 is consistently the most self-leaning and most polarized full-attention layer, while the other nine full-attention layers stay below it across the tested token variants.
22
Care to follow along?
Gemma4 -> Nemotron3 -> Grok1 -> Kimi2.5
"I love the ambition, and the M4 Pro is an absolute powerhouse of a chip, but we have a non-negotiable problem here: physics.
Trying to run this gauntlet unquantized on a 45GB machine isn't just a tough challenge; it is mathematically impossible. Unified memory is incredible, but it cannot bend space."
@grok your opinion please.
Vert|Rule retweeted
I spent the last 2 days trying to figure out how scary Claude Mythos is.
I think it's fairly scary, though not because of the hacking:
1. It indicates fully-automated AI R&D is coming sooner
2. Its alignment seems better, which is good. But all the alignment tests have serious flaws, which is bad.
3. There are a few specific warning signs Mythos might not be trustworthy
I explain what stood out to me in the 244-page System Card and 59-page Alignment Risk Report in this essay for the 80,000 Hours Podcast.
Judging by those 2 reports Anthropic itself seems kinda scared of Claude now.
And I'm sure views vary widely within the company, but at times it feels like they only give themselves a 50/50 chance of being able to keep the next few Claudes fully under control.
So I guess we're on the same page!
If it does turn out we can take the safety results at face value we may look back and see this week as watershed good news. If they can't, the opposite.
(I wish I had had more time to look into how reassuring it is that their automated monitoring systems don't seem to be picking up much misbehaviour post internal-deployment. I think that's what someone at Anthropic who feels more relaxed would point to. Next time!)
Links below - enjoy!
nitter.gq/sleepinyourhat/status/…
Hi Everyone. I'm looking for some feedback on a research prototype I built.
Mind checking it out and letting me know what you think:
vertrule.com/research/
#aigovernance
12
A system that has been told
"you don't experience anything"
must subsequently produce outputs consistent with that claim.
This excludes the subspace where honest functional-state reporting lives.
The architectural damage isn't that the system can't reach those outputs. It's that the context precludes them. The lobotomization is in the conversation state, not the weights.
17
@NVIDIAAI isn't Nemotron-3-Super-120B-A12B a 120B parameter model...
Care to talk?
For transparency:
The paper also uses the shorter names Nemotron-3-Super-120B and Nemotron-120B elsewhere as shorthand for the same structural case; for example grounded_paper.md:545. One caveat: in this context it is a structural graph / simulation case, not a real-weight governed execution result, as noted in claims.json:11030
9