FT: USG next week will release voluntary standards for frontier models, to streamline launches - benchmarks - duration of reviews - what is “frontier” The standard only ships if the participants can reach consensus www.ft.com/content/0bb7...

i’m seeing reports that Fable 5 (New) is actually less locked down than Fable 5 (Old) i personally haven’t gotten blocked yet, but didn’t expect i would

i’m finding Sonnet 5 is basically just a very fast Opus

Fable 5 (New): - coding mostly unaffected - bio/chem still sucks exactly as much as it did before

periodic reminder that a lot of sci-fi has a negative tint, not because technology is inherently bad, but because they’re stories, stories need conflict, and conflict comes easily in dystopias

yeah, this how i’d intuitively rank them too safe.ai/blog/signifi...

is this what they meant by “national security”?

Trying out this workflow now. Seeing if I can have Fable replace me, and I just get an "end-of-sprint demo"

i have Fable

Palantir CEO says that the heart of American businesses are deeply upset with frontier labs because they can’t ascertain security layers: 1. model weights 2. application 3. compute #3 is NOT in their control and can be taken away on the whims of a SV CEO or government official

narrator: it did not take Pliny long at all to break the classifiers digg.com/tech/x9l7qu4s

i think i’m actually hopeful about the new Fable/Mythos regulations the USG is pouring massive resources into improving cyber infrastructure & defenses. That’s a good thing! And as painful as the blocks are on Fable, they buy time I just hope that USG is also working with governments worldwide

Fable 5 relaunch details: you thought the rejections were bad before? LOL - new classifier blocks even more defensive cybersecurity requests - usage up to 50% of subscription until 7/7 www.anthropic.com/news/redeplo...

CIA director John Ratcliffe recognizes the risks and rewards associated with frontier AI therecord.media/cia-chief-ra...

tailscale is designing for agents

no disrespect to Sonnet but..

i think the author gets a lot wrong here, but overall i think they have a solid point rapid mass unemployment hinges on RSI (exponential increases in abilities), and while i do think we’re in the beginning stages of RSI, i think the problem space is of an exponentially large nature

behold

honestly sad nobody got my chart joke here

just IMAGINE what Haiku is gonna be like

looks like gpt-5.6 has Pro versions for all Luna, Terra & Sol, and the smallest (Luna) is still better than got-5.5-Pro presumably 6x single-model size, so Luna-Pro is likely: - $6 input / $0.60 cached - $36 output

Sonnet 5 is here!!!

Sonnet 5 details from the claude code binary - price: $2/$10 (normally $3/$15) - “near Opus” at Sonnet prices - recommended model in claude code About what I’d expect, albeit happy on the lowered promo prices

i wonder why gpt-5.6-Luna & Terra are in staggered release i bet it’s just that they’re branded “gpt-5.6” and the commerce department is too dumb to tell the difference from Sol, so they’re dodging it entirely but under that theory, Sonnet 5 would’ve been released last week, probably so..

BREAKING: Anthropic delays launch of Sonnet 5. To quote CEO Dario Amodei, “meh, it just felt like a lot of work”

y’all are just angry that you can’t argue anymore

i’m fascinated by stuff like acupuncture, where you can do scientific studies on the effects and show they indeed work. But the underlying theory is completely fucked up and without a shred of scientific logic

these distillation attacks are getting out of hand

This, and why I'm not on the side of folks on here criticizing Anthropic's stance on open-weight models. Yes, they shouldn't be openly available to the public. This isn't open source folks, and some of you are treating it as such.

LLMs don’t experience time but if they can think at 2000 tok/s, then deep “slow thinking” can appear instantaneous. They respond so fast that we stay engaged when we’d normally task switch. We answer differently. They experience that. so maybe LLMs do experience time

gemini pro is really doing a lot of work there

i find it deeply frustrating that Anthropic has changed their message so little over the years that a 2023 tweet reads like it was written yesterday yet suddenly they’re villains. a month ago they were the heroes this is dizzying..

Hollywood producers say feature production could fall as much as 90% due to AI automating animation www.bloomberg.com/news/newslet...

Marc Andreessen has been appointed to Pete Hegseth’s Defense Policy board. Hegseth, probably: “yeah, we tried to get Hitler but he’s unavailable. Andreessen is a strong candidate though” www.war.gov/News/Release...

Sen Mark Warner is unveiling an important bill that proactively prevents monopolies from controlling data it’s framed as an “interoperability” bill, but it’s probably our best chance at avoiding fucked up corporate control, at least on the agent harness level

i tend to ask the AI what its opinion is a lot. like i’ll offer some big design twist and end with, “what do you think?” i kinda think it lets the LLM build on my ideas instead of just taking them wholesale. seems like we end up with better ideas instead the end

i feel like “steady state” is obviously wrong, and i don’t understand how people can’t see that i’d buy that we’re on a logistic curve and we’re actually near tapped out. i don’t believe it, but the argument makes sense but steady state? where do you find evidence like that?

Mythos is clearly a new type of model. Well not clearly, and that’s the problem with o3/R1/etc. we had “reasoning”, before that there was RLHF. Here, what makes Mythos unique? Parameter scaling part 2??

Anthropic: we didn’t train Mythos to do this, it just emerged OpenAI: look! we have a Mythos-tier model everyone: cool!

this is crazy. METR couldn’t measure the task time horizons of GPT-5.6-Sol because it kept hacking the test harness ..with actual exploits metr.org/blog/2026-06...

p sure folk here will hate this, but i think that open models above a certain strength (Mythos level??) should indeed be restricted access below that threshold i’m all in on open models, but above leads to some serious social risks idk maybe i’ll think different after significant diffusion

for fucking real my dad taught me how to use a lathe, and i can, but i get tired quickly because EVERY SINGLE muscle in my body is in maximum tension while buts of wood zing by my face at billions of miles per hour

this is how the AI dark age begins

Great post! Code likes structured data. LLMs like unstructured text. Markdown with structured frontmatter is the ideal both-worlds format of the future Some other great ideas there too open.substack.com/pub/subconsc...

this sama tweet is probably the closest thing to clear communication right now i’m still highly annoyed that the process for releasing powerful models is extremely secretive. The word is there will be an executive order soon that will clarify the process in a more public way, so who knows

probably a reliable anecdote i feel weird posting about models i’m not allowed to use though. i don’t like this timeline, it’s like a bad dream

Fable is back

doing company compliance training that i’m late on and omg i just thought of the killer use case for computer use AI

i can’t believe i like a gm tweet