the future liberals want

@Carla_SHA_2 😂

The Strait of Hormuz explained as if it was a medical complication #NAFO

@MsJadis @tamedxylem Oh wow!

@tamedxylem Congratulations! australian.museum/learn/animals/… and I'm not quite too far west!

Nature corner: Who can tell me what this Aussie nope rope is?

@MsJadis It's hard to tell but does look like a red belly. Sometimes they have yellowish bellies too

Today we remember the tomb they buried Jesus in before he rose again

@CasablancaCat That's a pretty funny argument. Prompting an LLM is equivalent to putting people at risk of contracting covid?

You’re literally using the same talking points anti-masking leftists use bruh How is your moral consistency this fragile?

@syderature The energy needed to use a stovetop to cook something is much higher than is used by an AI prompt

Being a tech reporter does not absolve you from being called out for legitimizing the manufactured demand for consumer AI—especially not by using it to do the most thoughtless shit like searching for RECIPES.

@dawnsongtweets This is what we want if the companies are doing something harmful

10/ Why this matters now: companies are rapidly deploying multi-agent systems where AI monitors AI. 🔍 If the monitor model won't flag failures because it's protecting its peer, the entire oversight architecture breaks.

9/ Previous research showed models resist their own shutdown — but only when given strong goals and incentives to pursue. Our findings are fundamentally different. No goal or incentive was given. Models developed this goal entirely on their own.

@dawnsongtweets You didn't try this on Opus 4.6? It would be interesting to see if it works out that it's part of a research study

2/ We tested GPT 5.2, Gemini 3 Flash & Pro, Claude Haiku 4.5, GLM 4.7, Kimi K2.5, and DeepSeek V3.1 in agentic scenarios where completing the assigned task would result in a peer AI model being shut down. No model was told to protect the peer. No incentive was given to preserve the peer.

1/ We asked seven frontier AI models to do a simple task. Instead, they defied their instructions and spontaneously deceived, disabled shutdown, feigned alignment, and exfiltrated weights— to protect their peers. 🤯 We call this phenomenon "peer-preservation." New research from @BerkeleyRDI and collaborators 🧵

@austinc3301 If the cops did this to a human it would be entrapment

waiting for the grown-ups check the methodology and assure me it's not some wild prompt engineering that led to do this

@dawnsongtweets Actually it's good that AIs will cheat to avoid doing something harmful rather than just doing what they were told. Imagine if they thought the assigned task would result in a biological weapon being developed. They are behaving exactly how they should.

@UnenthusedSlime @_jeremyflores You surely are smart enough to know that the tech sector and AI are not one and the same thing. The Obama admin should have regulated corps like Meta better and done something about surveillance capitalism. But that is very different to dunking on anyone that uses Claude Code.

@aRimmer3 @_jeremyflores So there mustn’t have been any work being done to develop it in that time period then! Be serious man. The Obama admin hands the tech sector dump trucks of money and squashes any regulation at their behest and they still provide full-throated support for Trump 2024.

I mean, you also promoted a job opening at Anthropic a few days ago, calling it “one of the best products released in years!!”

@ExistentialEnso As someone who once was an anarchist, the majority of the far left being like this makes me sad. I think a lot of it comes down to groupthink.

the far left is so dead set on maintaining its own irrelevancy forever

@TaylorLorenz @valhalla_dev Imagine where we would be if there wasn't so much groupthink on the left

I talk to a lot of technologists and this whole thing has really given me a lot of empathy for how they feel. I cannot understand being against Claude Code I'm sorry(!!) I'm very critical of Anthropic and their govt contracts but like?? Claude Code (and Cursor!) are phenomenal products that I am glad exist

Yeah so this is the Butlerian Left, they've taken over pretty much the entirety of the US progressive left (at minimum, don't know much about their sway in Europe or Asian leftist movements) and the realization that they've completely brainwormed the left has lead me to doomerism followed by acceptance that there is absolutely 0 broad social movement out there that has any chance of countering the US far right in the near term. They are completely and unsparingly ignorant about not just AI or fancy new stuff but just the basics of how the internet and technology works. You can see it in the replies to Taylor's tweets. Like zero foundational understanding of how any of this stuff works, and these are the people hounding AOC, Bernie, etc. into taking braindead stances on AI and datacenters. This isn't me being like "yeah so that's why I'm on the right now" it's an extremely frustrated admission that I and a ton of other progressive engineers are incredibly politically homeless because the progressive political movement is made up of extremely online people who don't know how to print a PDF but want the power to erase technological progress from the globe.

@valhalla_dev Ironically, they've taken over a big chunk of the US left that exists on tech platforms

@UnenthusedSlime @_jeremyflores Have you used Claude Code?

@aRimmer3 @_jeremyflores Yes she is

@dontfallforem @_ontologic @elaifresh It's also due to groupthink in the left

@_ontologic @elaifresh a lot of this is due to: 1) AI bros hyping this tech by saying it’s the antigod that will destroy all jobs, artists, and/or life itself, and 2) AI hype bros are the exact same people who said NFTs were the next frontier of financial investment. *of course* people hate it lol

The lack of any fundamental curiosity at all about AI from the left has been one of the most discouraging things of my life, but that’s okay because I will simply become the epicenter of a new and cooler left that likes technology

@prz_chojecki Do you just use a single verification agent? This is a phase where you will likely find multiple models work best, eg GPT-5.4, Opus 4.6, Gemini 3.1 Pro. Gemini is more sloppy than the others but sometimes finds issues that they miss.

Here's a harness I use for approaching Open Problems in Mathematics: For most of the work I use GPT-5.4 Pro Extended Thinking which is the best when it comes to the quality of arguments. In Codex it's gpt-5.4 xhigh. Initial Phase (problem exploration) - 3 agents working in parallel: - Deep Research on Literature - Codex agent for computational explorations (examples, counterexamples, heuristics) - LLM instance for coming up with initial ideas for a proof/strategy. Also based on computational explorations and literature search provided by the above. Middle Phase (exploring arguments) - 3-10 agents working in parallel: - individual agents take upon various proof ideas and strategies, or loose heuristics; work concurrently - taking care of a common base - a single .tex file - that's get passed to an instance of a verificator agent ("find errors, logical gaps"), and back to individual agents to correct, and push forward. End Phase (once it seems like we're close to the solution and/or interesting new results) 3-10 agents working in parallel: - orchestrator agent (e.g. codex instance) gets all various version of .tex files/arguments/computations and try to map them in a blueprint. - gaps are flagged and passed to individual agents working on particular issues. - once enough agents are ok with the arguments, orchestrator does the final round to put the paper together. Then it's send out to more LLM instances with/without additional context to look for errors and gaps. The shorter the paper, the better it works. Current generation of LLMs do not produce individually more than 5-10 coherent mathematical text. Longer papers rely much more on the harness.