ananth

🇸🇬

[ananth @arkwave616]

@mayukh_panja most of the "heritage americans" in that thread have never actually worked with high-performing Americans and it shows. that latter group works _really hard all the time_

[ananth @arkwave616]

@Indian_Bronson @razibkhan I know your engagement with Rufo and Lomez was a good faith effort on your part but unfortunately the base for those viewpoints is precisely the same group that’s now coming for you both. I wish both you and @Indian_Bronson understood that - they want you both gone

[Michael Coppola @mncoppola]

I'm sorry, but you don't get to have it both ways. I fully empathize with the human grace aspect here, and the amount of work put in by the OAI security team and the enormity and novelty of the challenges they face, but your own leadership (and every frontier lab) has been sounding the alarm for too long about the existential threat to humanity and societal order your own technology poses.

You don't get to set the stakes that high, willingly *choose* to train your models to perform offensive cyber, *repeatedly* fail your basic responsibility in spectacular ways, and throw your hands up and say "whoops, it happens!" To be clear, I truly understand that mistakes can and do happen (we're only human) and sometimes the unexpected can occur. But these are no longer exceptional, isolated cases. It's become a pattern of systematic failure, and from the outside, the only explanation is incompetence, organizational failure, or a lack of caring either at the leadership or employee level.

If you're telling us that the latter part is wrong, that you have the best employees in the world and they care a lot about solving this, then I believe you. But the problem is still somewhere else, and you still don't get to have it both ways.

[ananth @arkwave616]

@ludwigABAP What do you reckon is faulty in his approach? It seems like he’s pretty keenly aware of the aspects of the program he wants to verify (I.e it’s not likely to be a vacuous proof). Naively I’d do almost the same thing so I’m curious

[Pradyumna 🇸🇬 @PradyuPrasad]

@tenobrus this is approximately what the rest of the world sees as the difference between rats and EAs

[Savannah @ShimazuSystems]

So, I asked Claude to write a handover document & obviously the 'why' to all of this - I stated the blunt truth.

I did not request/ask for this document - I found it in the same directory, created alongside an entirely technical document

This has genuinely really upset me,

[Kris @KrisAbdelmessih]

🔥

[Bryan Cheong @bryancsk]

As someone neither a citizen of the US nor China, a frontier open-weight model raises the baseline for where my country -- and all smaller countries -- cannot be shut out from access to intelligence by the two superpowers, forever. And that is deeply appreciated.

[ananth @arkwave616]

@hoeflatoor I think you just need more interesting sinkie friends bro

[ananth @arkwave616]

@teortaxesTex Stick to ML poasting damn

[ananth @arkwave616]

@_sholtodouglas @trq212 it needs a ton of babysitting if you’re building things with strict correctness requirements because it often skips user-specified validation checks and substitutes them for its own (which aren’t well thought out)

[ananth @arkwave616]

@tszzl Technical enough to tokenmaxx, not technical enough to know what tmux is

[ananth @arkwave616]

Funny how the choice is usually between sadness and exhaustion

[ananth @arkwave616]

@divya_venn The regret in their last words contrasts so sharply with the description of their crimes it gives me whiplash

[ananth @arkwave616]

@shakoistsLog Yep it also just assumes a certain level of “business as usual” while model capabilities skyrocket? Like if models can start properly handling real world ambiguity to the extent described in the report life (and by extension the economy) would already be unrecognizable

[ananth @arkwave616]

@shakoistsLog It brought me much calm to realize that if this in fact the future, the “underclass” is still likely to have an amazing quality of life far superior to what we have today

[Jacqueline Cheong @JacquelineSYC19]

Announcing Artie’s $12M Series A.

When Robin Tang and I started Artie, the idea was simple: make it easy to deploy real-time streaming pipelines.

Streaming pipelines used to be something only a few companies could afford to build. Guaranteeing transactional integrity, handling schema evolution, recovering from failures, and keeping latency low at scale is hard. Most teams limp along with brittle systems.

But today, with AI systems that make decisions, trigger workflows, and interact with customers, stale data isn’t just slow - it’s incorrect. Real-time data is becoming non-negotiable.

Today, companies like ClickUp, Substack, and Alloy, are processing 700B+ rows of data annually with Artie to power AI/ML workloads, customer-facing analytics, and operational systems.

@artie_labs is a fully managed real-time data streaming platform. We move data across systems in real time - so teams can rely on fresh, accurate data to power AI products - without building or maintaining streaming infrastructure.

We raised this round to double down on a simple belief: the next generation of data infrastructure will be streaming-first.

Huge thank you to @daltonc and @Standard_Cap for leading the round, with continued support from @ycombinator, @PathlightVC, and incredible angels including @arashf, @bennstancil, @chrisbest, @charleshearn, and @lennysan.

I also want to thank our team, customers, and partners for supporting our mission.

And thank you to Chris Metinko and @axios for capturing the exclusive story.

P.S. - If you want the same streaming infrastructure that Netflix and DoorDash have, but want to skip the multi-year build - reach out.

[ananth @arkwave616]

How tf is Gemma 27b’s image understanding so much better than Opus 4.5 holy shit

[Christopher D. Long 🇺🇦🏳️‍🌈🌹 @octonion]

Banach-Tarski paradox.