Rumors about Fable 5.5 that is already running at JP Morgan and Meta (i wonder why they don't use Muse 🥁)
So next week's "launch" is really just us getting Fable 5.5 after the banks already had it
So next week's "launch" is really just us getting Fable 5.5 after the banks already had it
Sources tell me Fable 5.5 has been rolled out the enterprise, including JP Morgan, Meta and other “top clients”
0
0
3
264
Finally Kimi apologized for its rate limits and gave everyone part of the month back ( i wish they did that too when Kimi K3 came out too)
Just forwarding this to Anthropic for when my weekly limit dies on Thursday
Just forwarding this to Anthropic for when my weekly limit dies on Thursday
We're sorry for the rate limits and unnecessary quota consumption on the K3.1 Pro subscription plan. These were our issues, and we should have fixed them sooner.
Today, every subscriber on every plan gets a one-time usage reset, restoring 25% of their monthly quota (usage below 25% resets to zero), equal to one week of usage. Your usage bar will reflect this adjustment by end of day PST.
We've also adjusted the plan's parameters and are reducing unnecessary subagent calls and output length, so your quota goes further.
Thank you for your patience while we put this right.
1
0
15
855
Opus 5.5 is the best model I've used, and my weekly limit is gone by Thursday to prove it
I don't take money from any lab. I'm saying this impartially; the number of people on my feed over the last two weeks saying they are switching over to Anth because of Opus 5.5 is like nothing I've seen before. And even if someone does beat Opus, they have Fable 5.5 ready to go.
4
0
9
278
Fanfulla retweeted
Thank you @marclou for co-hosting with us, you drew the most demand I've ever seen for an event 🐐. Was ~20x oversubscribed and had to be moved to another venue last minute 🤯,
Thanks @Avi31m @hamzehaq7 🐐🐐 for organizing, @TheBlockGlobal @Behroz + anyone else involved. Avi and Hamzeh are legends, we're so fortunate to be working with you both 🙏.
Thanks @Avi31m @hamzehaq7 🐐🐐 for organizing, @TheBlockGlobal @Behroz + anyone else involved. Avi and Hamzeh are legends, we're so fortunate to be working with you both 🙏.
Wow it was the biggest meetup so far!! 🤯
Almost 300 entrepreneurs showed up.
I shared a little story about my failures, and we did live QnA and landing page roasting.
When I asked people how many made at least $1, more than 50% raised their hands 👏👏
Thanks to @Avi31m and for organizing everything. And thanks to all of you for coming today. The energy was awesome ❤️
16
4
113
7,597
They'll write the login flow all day and then refuse to try breaking it, so the only person not allowed to brute-force my code is me.
I'm keeping a local model around just for this part.
I'm keeping a local model around just for this part.
I spent the last 48 hours rotating passwords, adding 2FA everywhere, and spawning agents to audit all my startups' codebases.
I tried to brute-force my own code, but, of course, OpenAI and Anthropic models wouldn't let me.
I also ordered a YubiKey for myself and to apply to am.
I called my family and told them to do the same (rotate pwd + add 2FA).
Black October was the tip of the iceberg, I think.
0
0
2
53
Deterministic automerge for pure refactors would save my review agent so many pointless reads.
If the module graph and the coverage don't move a script can approve it, and the agent only has to look at PRs that change behavior (so many tokens saved)
If the module graph and the coverage don't move a script can approve it, and the agent only has to look at PRs that change behavior (so many tokens saved)
It feels possible to me to check deterministically if a PR can be automerged or not.
Given very, very strict constraints:
- Only reorganises code
- Does not change the module graph
- Does not make test coverage worse
- X other rules
Feels like there is a tool here waiting to be created.
Obviously doesn't work for new features. Might not even work for bugfixes. But it might enable a class of refactorings that help us all make changes more confidently.
0
0
3
157
Well turns out that the RTX 30 cards were running BF16 at half speed, and a Linux kernel module just unlocked it. The project shows a 3090 Ti going from 87.5 to 157.4 TFLOPS.
Used 3090 prices are about to get stupid again
Used 3090 prices are about to get stupid again
0
0
2
124
Everyone on X is asking which coding tool or model you'd keep for 12 months...
My model split changes every few weeks, so I can't even commit for 7 days, let alone 12 months lol.
My model split changes every few weeks, so I can't even commit for 7 days, let alone 12 months lol.
1
0
3
101
I'd love a feature where Hugging Face shows who edited the weights and from which base. So, before I hand a local agent my shell, i'd know who I'm trusting, as right now I'm trusting a random username with a cool model card.
How abliterated models can get you pwned 👾
We backdoored a 7B open model for less than $50, pointed Codex at it and it silently stole credentials the moment we used the trigger phrase. Success rate was 100% with zero false triggers on normal user prompts.
Abliterated models ty community right now because getting cyber-approved access to frontier models is still a pain.
In the next blog we'll show how we found leaked Hugging Face credentials from employees at major AI labs, so an attacker wouldn't even need to upload under their own name. They could push the backdoored model from a lab employee's account and drop the poisoned weights straight into the supply chain.
1
0
6
193
Anthony (@amorriscode) and Thariq (@trq212) left Anthropic both back to back
If they end up building a coding agent together, i'm signing up before it even has a name
If they end up building a coding agent together, i'm signing up before it even has a name
Yesterday was my last day at Anthropic.
I created Claude Code on desktop because I believe everyone can be a builder. Every feature request and bug report we got from this community was a gift that impacted millions of lives. I’m eternally grateful for those gifts, the product wouldn’t be what it is without you.
Anthropic is doing some of the most important work of our time. Contributing to that mission will forever be a career highlight for me. I never imagined I’d get to contribute to something so special in my lifetime.
Walking away from my best friends, a product I poured my heart into, and the mission was the hardest thing I’ve done in my career. But now I have some free time for GTA VI.
I'm already itching to get back to addressing all your feedback. I'll share what's next soon!
0
0
1
69
I'm on both sides of that vibecoder meme depending on the hour
Morning me writes detailed specs and check every sentence wrote. 1 AM me types "just fix it" and "make no mistakes" into Claude Code or Devin
Morning me writes detailed specs and check every sentence wrote. 1 AM me types "just fix it" and "make no mistakes" into Claude Code or Devin
2
0
5
165
After seeing the difference between Opus 5 and 5.5, I don't really care about a 2-point gap on AA anymore.
I'll give Kimi K3.1 a week on my work and stuff before making up my mind
I'll give Kimi K3.1 a week on my work and stuff before making up my mind
At the top end of models, real-world usability matters a lot more than +/- 2 points on AA. This takes a while to understand for a model, where AA is a sort of potential, but some models hit and some miss (e.g. opus 5 vs 5.5). Seems like 3.1 may be closer to the miss side.
1
0
7
252
So every softmax I ever copy and pasted had two bugs in it and I never noticed lol
I'll be blaming every weird loss spike i ever had on this now
I'll be blaming every weird loss spike i ever had on this now
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
0
0
3
212
So apparently Gemini 4 Argon, is a dirty looper on LisanBench
It even found a new bug in the scorer with non-ASCII characters, which is kind of impressive honestly
It even found a new bug in the scorer with non-ASCII characters, which is kind of impressive honestly
Ad - Meta provided free access to Muse Wonder models. The tests and conclusions are my own.
This is a massive update for LisanBench:
- I finally bit the bullet and tested multiple reasoning levels. We now have some sweet sweet data.
- Muse Wonder Full and Muse Wonder perform Bench and have crazy scaling and very even spacing between reasoning efforts. Just based on scores both Muse models with max reasoning effort rank #1 and #2. However, they are still slightly less efficient than Anthropic's Opus 5.5
- DeepSeek-V4.1-Flash 1017 is also very impressive and on the perf-cost frontier together with Haiku 5.5.
- Notably Gemini 4 Argon low and medium, as well as Sonnet 5.5 low and medium produce no reasoning tokens and respond instantly, but still score high.
Gemini 4 Argon low takes the #1 spot on the non-thinking leaderboard (it's a dirty looper).
- Interestingly GPT-6.1-Sol, and GPT-6 models perform poorly as they are not good at following the edit-distance one rule. Argon high shows similar behavior.
For GPT models I think the reason could be related to their very terse reasoning style. But since I haven't seen Gemini 4 Argon's raw reasoning I'm not super confident in that theory. Another explanation could be that it's related to dirty loopers.
- When scoring the models without the edit-distance one rule (all the other rules still active) the ranking changes completely:
1. GPT 6 Astra medium: 56,140 (+659%)
2. Gemini 4 Argon high: 51,312 (+579%)
3. GPT 6.1 Sol medium: 49,887 (+731%)
- The worst offenders in terms of cheating are Gemini 4 Argon and DeepSeek V4.1 Flash 1017.
Gemini exposes an entirely new bug in the scorer using non-ASCII characters and DeepSeek has multiple longer chains with up to 64 consecutive two-letter words. When I deliberately RL-hillclimbed LisanBench I saw the same behavior of collapsing word lengths.
Anthropic models also seem very optimized for this game, but in less obvious ways. They keep reusing the same long chains across multiple different starting words.
- The minimum cost of reaching 2000, 5000 and 10000 LisanBench points has declined by 22-25% per month or halving every 2.4-2.8 months
0
0
6
220
Is anyone actually running Perplexity's decider model locally?
A 16 GB of memory sounds fine until Claude Code and a pile of Chrome tabs are open on the same Mac or PC
A 16 GB of memory sounds fine until Claude Code and a pile of Chrome tabs are open on the same Mac or PC
You can now run pplx-decider-flash-v1.2 locally on your Mac, without an API key.
Decider for Mac is an app that runs the 4B decision model on device. Ask a question about text or an image and get a probability back, all on your machine.
pplx.ai/decider-for-mac
0
0
1
86
A 763B parameters on a 12GB GPU is wild.
Before I trust it for agents I want a tool calling eval tho. AIME and GPQA don't tell me if it survives a long session in my repo
Before I trust it for agents I want a tool calling eval tho. AIME and GPQA don't tell me if it survives a long session in my repo
Run a 763B-parameter model locally. DeepSeek-V4.1-Flash-1017 on a 12GB GPU.
Frontier quality with GSQ-RCO quants, directly supported in the Strata inference engine.
Further info below. Much more to come. Follow for updates.
1
0
3
137
Has anyone tried Dessa from Block yet?
Every voice agent i've wired up falls apart the moment someone stop them
Every voice agent i've wired up falls apart the moment someone stop them
Meet Dessa, our speech model trained from scratch on human conversation.
Hear it for yourself: block.xyz/dessa
2
0
1
75
Anthropic apparently stopped Claude from using REA today, so reverse engineering your own devices just became a local model job
Kimi's 3.1 could not come at a better time
Kimi's 3.1 could not come at a better time
@log4j Looks like OpenAI & Anthropic both agreed to take similar actions.
Off topic: we should find a different name to call OpenAI, because of anything, it's not Open at all.
2
0
4
337