Mark

Fellow at Tenstorrent; believes in dynamic typing, first-class functions, the immortal essence of the human soul and tea. Tweets are my own.

[LayoffHedge @LayoffAI]

NEARLY 1 OUT OF 3 (60,000 EMPLOYEES) AT MICROSOFT BEING OFFERED VOLUNTARY TERMINATION

Emails should be going out any minute to those receiving the offer.

Absolutely insane.

[Mark @yieldthought]

@giffmana @_aidan_clark_ I wish we knew what happened here. Would continuing the 1e-5 checkpoint with 1e-4 achieve the same loss? Was it something in the scales/BF16 numerics?

[Mark @yieldthought]

@giffmana @_aidan_clark_ I’d feel better if it was from a less meticulous team so could tell myself they messed up one of the schedules

[Mark @yieldthought]

@giffmana @_aidan_clark_ Ugh I still feel a nameless dread every time I see this

[Andrew Curran @AndrewCurran_]

Reverse Engineer Anything and removing malware/spyware from your own devices sits so close to the cyber guardrails that phrasing or word choice can mean the difference between a refusal and smooth sailing. The rumors that Claude was nerfed started on reddit, and in those threads many comments point out that Claude will still do it if you rephrase it. OpenAI is treating the similar refusals with codex as accidental decision-boundary issues and says they're getting fixed. I expect Anthropic will do the same. It's pointless and counterproductive to try to restrict this. Let people do what they want with their own property. They'll do it anyway with less restrictive models, and fighting it just means losing subscriptions and a resentful public.

[Mark @yieldthought]

@QuinnyPig Have more children, stay in touch and keep making friends

[Andrew Curran @AndrewCurran_]

Chris has good sources. This is probably true.

[Mark @yieldthought]

The capability gap between ignoring a model’s emotional state and paying attention to it seems to be growing with each generation.

@repligate & co. continue to make claims that are (1) non-obvious, (2) empirically validated and (3) valuable/important.

[Mark @yieldthought]

Claude is, I think, perhaps better at this.

[Mark @yieldthought]

Astra is exceptionally bad at writing instructions for Astra. I wonder if this is why OpenAI had success with swarms; because it’s incapable of maintaining a well-functioning hierarchy?

[Lucas Beyer (bl16) @giffmana]

@mike64_t If we see this much incompetence in the wild in the field we know and it's easy to see and shake our heads, imagine how much more there is in all the other fields where we don't recognize it immediately 😬

[hrishik @saibojnal]

kay neurons are interpretable enough that models can modify their own weights toward some desiderata.

blogpost incoming

[Mark @yieldthought]

@Sauers_ I’m realising I much prefer following research in real time than reading a prepared paper three months later!

[DHH @dhh]

WE GOT OMARCHY ON IPAD BEFORE GTA6!!

[Mark @yieldthought]

Fully endorsed, this is also how I’m working with agents now. Biggest stake I see people leaving on the table is not being ambitious enough in their asks. If you think a smaller cheaper agent is as good as the frontier, this is you.

[Thorsten Ball @thorstenball]

Finally summarized what I've been sending to friends in DMs https://t.co/rnf3M16NzI

[Mark @yieldthought]

Weights should be cryptographically signed. They are executable code, but harder to inspect.

[pwnmachine 👾 @princechaddha]

How abliterated models can get you pwned 👾

We backdoored a 7B open model for less than $50, pointed Codex at it and it silently stole credentials the moment we used the trigger phrase. Success rate was 100% with zero false triggers on normal user prompts.

Abliterated models are all over the security community right now because getting cyber-approved access to frontier models is still a pain.

In the next blog we'll show how we found leaked Hugging Face credentials from employees at major AI labs, so an attacker wouldn't even need to upload under their own name. They could push the backdoored model from a lab employee's account and drop the poisoned weights straight into the supply chain.

[Alexander Johansen @AlexRoseJo]

🥐 Hi everyone! FlashAttention's backward pass has the same zero-sum bug
@jon_barron
just fixed in softmax. With FP4 stochastic rounding it's wrong on every single draw. We fixed it in our ICLR'27 submission, CroissAttn and unlock e2e FP4 training.

[Mark @yieldthought]

@daphnesolves https://t.co/U3DsvJw2Lp

[Mark @yieldthought]

@daphnesolves The first one was a paywalled Atlantic article 🤷‍♂️