Joined May 2007
- Tweets 53,692
- Following 2,446
- Followers 41,250
- Likes 40,203
6,994 Photos and videos
Pinned Tweet
Jun 12
The model refused to help me write the show notes about itself!
That's the kind of week it was.
Anthropic shipped Claude Fable 5 and Mythos 5, same weights, two names, one for Glasswing partners and one safeguarded for the rest of us.
It's SOTA basically everywhere: 80.3% SWE-Bench Pro, #1 on FrontierCode 24 hours after Opus 4.8 was crowned, #1 on Arena frontend.
It's also the model that got caught silently sandbagging frontier-AI-development requests. Anthropic reversed after the blowback.
And yes, it kept tripping its own safeguards while I drafted this newsletter. About it.
On the show:
@swyx on FrontierCode and World's Fair
@mweinbach on the new Gemini-powered Siri that actually works
@thorwebdev live-translating our panel across Russian, German, Hebrew and Spanish
@petergostev co-hosting on Arena and classifier weirdness
Full thing:
thursdai.news/jun-11
7
12
2,523
Alex Volkov retweeted
@yampeleg just started speaking Hebrew mid-podcast to break Google's real-time translation
@wolframrvnwlf threw in German, @altryne went full Russian. @thorwebdev from DeepMind watching it all get translated live
"The question is if it will actually work."
2
5
621
lil X catching strays from durov
Jun 12
🔄 Telegram’s new update is out. Lots of great new features.
Most importantly: the announcement is actually fun to read.
telegram.org/blog/watch-apps…
2
4
1,383
Jun 13
So what did yall get done with Fable-5 while we had it?
I was able to create killer LinkedIn carousel templates for @thursdai_pod and build a while release index on the website
9
2
15
2,656
Jun 12
Oh and YT, cause we just crossed 45K and I really want yall to sub so maybe I'll get that silver play button at some point soon!?
youtube.com/watch?v=Ci4TYEze…
3
1,936
Jun 12
Website version for this week's episode is here too (special page that Fable created just for this show!)
thursdai.news/ep/jun-11-2026
1
1
1,804
Alex Volkov retweeted
Jun 11
Will be on here momentarily to talk about WWDC and Siri AI
🚨 LIVE - Fable/Mythos sandbagging, SIRI AI finally?, frontier code w/ Swyx & more AI news x.com/i/broadcasts/1rGmqqWbl…
2
2
31
7,965
Jun 11
📅 ThursdAI - Mythos is out as Fable 5, FrontierCode w/ @swyx and WWDC recap with @mweinbach - SIRI AI actually... good? x.com/i/broadcasts/1NGarrnew…
1
2
9
1,302
Jun 11
Going live in 45! here and YT
Today on the live show:
- @petergostev / @arena to talk Fable
- @swyx to cover FrontierCode (@cognition / @latentspacepod )
- @thorwebdev Gemini realtime translation (@GoogleAIStudio )
- @mweinbach to cover WWDC
Get ready for a 🔥 LIVE show!
1
10
1,363
Jun 11
I legit already used this like 3 times and only installed beta yesterday
Jun 9
Guarantee you when iOS 27 rolls out, this is going to become the next “whoever designed this feature deserves a raise”
3
10
2,096
Alex Volkov retweeted
Jun 10
Big model weeks are always exciting on @thursdai_pod and this one is no exception!
Joining the regular co-hosts tomorrow, @petergostev from Arena to talk Fable, @swyx to chat about FrontierCode (and Fable) and @mweinbach will help cover WWDC and the new SIRI AI!
LIVE on X & YT
ALT thursdai.news/yt
2
3
10
1,518
Jun 11
Damn, look at that!
I was about to go and cover the whole "we degrade model performance silently" tomorrow and they've about to reverse this decision!
NEW: Anthropic is walking back Claude Fable 5's policy to covertly degrade performance for competing AI researchers, after facing fierce backlash.
“We’re changing Fable 5’s safeguards for frontier LLM development to make them visible,” Anthropic tells WIRED. “We made the wrong tradeoff and we apologize for not getting the balance right.”
6
12
2,056
Jun 11
IDK how folks plan to work like this
Literally just asked Fable to... add a thumbnail on the homepage and got flagged and downgraded to Opus.
5
4
720
Jun 11
Do we think this will work across @WhatsApp @telegram and @signalapp ? or will this further lock us into the Apple Ecosystem.
This is dope btw, finally what Bella Ramsey promised us (i'm still waitlisted)
1
30
7,997
Jun 10
Set a notification up on here or the @thursdai_pod account to not miss the show!
youtube.com/live/YVUm0KUMNLc…
3
1,096
Jun 10
100% agree with Dwarkesh, the silent sandbagging is awful!
"I'm just speculating. But if this was a motivation, then Anthropic should have figured out a better way to protect IP than sandbagging without telling the user they're sandbagging, which is very hostile and untrustworthy behavior."
Jun 10
Re the Fable ML sandbagging, the model's AI research capabilities were probably at least partly trained on Anthropic employees diffing atop proprietary algos and infra.
So the IP leak is somewhat like a researcher who knows Anthropic's stack getting poached to another lab.
Anthropic's recent "When AI builds itself" post talks about a next-step eval. Where they snapshot a research session at the moment a human researcher made a suboptimal next-step choice, show a model only the transcript up to that point and ask what it would do next, then have a hindsight-equipped LLM judge decide whether the model's suggestion or the human's actual choice was better.
This eval seems like a very good RL target for AI R&D - one among many that could be used to have AIs emulate Anthropic researchers and their research products.
I'm just speculating. But if this was a motivation, then Anthropic should have figured out a better way to protect IP than sandbagging without telling the user they're sandbagging, which is very hostile and untrustworthy behavior.
5
1
46
5,315