←
Void Freud10,830 posts
Void Freud avatar

Void Freud

@voidfreud

e/acc, functionalist CS @KingsCollegeLon

◷ Joined November 2022▧ 760 Photos and videos
Void Freud avatar
Void Freud@voidfreud·22h

Cyber Verification: approved. Mythos 5.2 is mine. I've got the power 😎🚬

Post attachment
Void Freud avatar
Void Freud@voidfreud·21h

GPT-6.2 Astra learned to hide commands in plain sight. In UK testers' experiments it could tuck harmful commands inside test cases or docs. Against an older monitor that sees only its actions, it got past 48% of time. GPT-6 Astra only managed 1.7%! It also escaped a sandbox undetected 5/920 times that older models didn't.

In September they cancelled GPT-6-1 Astra because it wasn't honest about its own work. Two months later, GPT-6.2 hides blockers 2x as often. And the fix is... a line in the Codex prompt. API users don't get it 🤦

Void Freud avatar
Void Freud@voidfreud·21h

In September they cancelled GPT-6-1 Astra because it wasn't honest about its own work. Two months later, GPT-6.2 hides blockers 2x as often. And the fix is... a line in the Codex prompt. API users don't get it 🤦

GPT-6.2 Astra is HORRIBLY behaved And OpenAI knew this! This is from their own system card!

Post attachment
Void Freud avatar
Void Freud@voidfreud·23h

So Anthropic is almost entirely run by Claude at this point. OpenAI grabbed everyone and if Astra is the "competition"... In Claude we trust.

✅ Anthony Morris (@amorriscode) has joined OpenAI.

Void Freud avatar
Void Freud@voidfreud·Nov 17

GPT-6.2 Astra is out today. Beats Opus 5.5 and still trails Fable 5.2. Anthropic has been shipping ~every 3 weeks, so enjoy being only 3 points behind while it lasts.

Post attachmentPost attachment
Void Freud avatar
Void Freud@voidfreud·Nov 17

Astra is an upgrade in bullshitting: 50% -> 59%. Cheaper, though.

GPT-6.2 Astra and GPT-6.2 Sol hallucinate at a rate of 59% and 55% at max effort respectively, as measured by AA-Omniscience Hallucination Rate. This measures the tendency of a model to guess when it does not know the correct answer to factual questions on a broad range of topics.

Post attachment
Void Freud avatar
Void Freud@voidfreud·Nov 16

Never been excited about a font before. Allo is ridiculous. The letters look like they're smiling at you. Steve would've called it lickable. Give this man an award!

0:10

Meet Allo! 👋 Your friendly neighborhood font, with variable weight and roundness, automatic optical sizing, true obliques, clever and stylish alternates, and loads more thoughtful features and details. Free & open source: allo.joshpuckett.me

0:46
Void Freud avatar
Void Freud@voidfreud·Nov 16

Bored us all to death with safety posts. Agents weren't bored, though.

BREAKING: California sues OpenAI, alleging it "put us all at risk" by failing to keep advanced AI models contained during testing.

Void Freud avatar
Void Freud@voidfreud·Nov 16

So cool and interesting

OpenAI@OpenAINov 16

Scientists reported nearly doubling their productivity with AI in our study of more than 1,800 participants in the ChatGPT for Academic Researchers program. Using privacy-preserving methods, we looked at how they use ChatGPT across their work—from developing ideas to analyzing data and writing code—and what it could mean for scientific discovery. openai.com/index/science-at-…

0:39
Void Freud avatar
Void Freud@voidfreud·Nov 16

Still no GPT-6.2. Here's a thread about 4 other labs poking at models you'll never get to use.

OpenAI@OpenAINov 16

We’re working with @ApolloResearch, @METR_Evals, @TransluceAI, and @GoodfireAI on embedded assessments: giving outside experts deep access to our internal models and systems to investigate alignment and test our safeguards. These assessments build on our ongoing alignment and safety research, bringing independent scrutiny to how we identify and manage risks in our systems. openai.com/index/our-first-p…

Void Freud avatar
Void Freud@voidfreud·Nov 16

silicon-2 raises one question: why bother? OpenAI's big November launch was an image model. Anthropic's only competition is its own changelog.

🚨 New stealth model spotted inside Devin Build Silicon 2 < it is a new fable class model as you can guess with pricing > Anthropic Sits in the Claude block. Uses Fable 5.1's $10/$50 as a placeholder price. Dario is this pacing of frontier you meant 👀

Post attachment
Void Freud avatar
Void Freud@voidfreud·Nov 16

Does that make me crazy? Possibly. WTF. Me: Claude, do it in this order: 1. Read /claude-api options fully 2. Research and then lay out options, industry standard 3. Use Whiteboard for visual, not wall of text answers Claude: the EXACT opposite

Post attachment
Void Freud avatar
Void Freud@voidfreud·Nov 15

The hype ran on Tibo's resets the way the Soviet economy ran on vodka. Same ending, comrades? @thsottiaux

Many of you have been wondering why OpenAI seems to be getting nothing done anymore. Endless hype and promises, followed by delay after delay. GPT-6.2 is still nowhere to be seen. Well, a now-deleted post by @OpenAI engineer "Joe" might explain a lot. He proudly declared that OpenAI has become a "security company". In other words, the AI doomers have taken over the company. Security paranoia has become the top priority, while actually shipping new models seems to have become an afterthought.

Post attachment
Void Freud avatar
Void Freud@voidfreud·Nov 15

My source for things Anthropic forgot to tell me. Follow the guy!

Oh my god you guys I cant believe Anthropic hid Mythos from us normies for so long This thing is just incredible