Angehefteter Tweet
efe
2.1K posts
efe
@extliqprovider
druckenmiller is my GOAT what the fuck is this world? you can't hedge a worldview
Beigetreten Şubat 2025
539 Folgt83 Follower
efe retweetet
@AgustinLebron3 @AnthropicAI this is just consistent with their prior behaviour and thinking
if they had the mandate then they are not losing it because they are the same anthropic
English
Again, they're not nerfing ML research by refusing requests. Instead, it quietly sabotages users by lying to them.
@AnthropicAI is steadily losing the Mandate of God.
Jeremy Howard@jeremyphoward
@karpathy This is not a day for celebrating, Andrej. It's a very dark and very sad day, and the damage may be impossible to undo.
English
@bubbleboi @zephyr_z9 btw elon said anthropic is good people a month ago 😂😂
English
Have canceled my team subscription for Claude Pro.
Idc how good that model is, it’s not good enough for me to support people who actively stifle innovation and gate keep knowledge that they didn’t even create.
English
All oai folks now have to do is to release the big boy they have without sandbagging and anthropic will start hemorrhaging market share right before ipo
English
@gbrl_dick dario was honest from the beginning that there shouldnt be any open source ai
English
@LeonHowqua @DevelopmentsAI @teortaxesTex so you think AGI is achieveable just by improving coding + synth data and nothing else needed?
English
@extliqprovider @DevelopmentsAI @teortaxesTex At this current point in time, no fancy new math/science is really needed to improve the LLMs. It's just more efficient training code, architectural experiments, scaffolding, data generation, RL environment building etc, which is achieved with better coding capabilities
English
I really don't think OpenAI is going to let this slide.
I've been saying it for a long time, the real inflection was when they reached 5.2. I have no clear insight on what they currently have internally, but if they haven't made a Mythos/Fable yet, it was *a choice*.
Andrew Curran@AndrewCurran_
The internal boost from Mythos-assisted development since February is just too big. Anthropic is pulling away from the pack for the first time, and at the same time they are also speeding up. The race legitimately feels like it is changing for the first time in years.
English
@LeonHowqua @DevelopmentsAI @teortaxesTex isnt coding just a tool to implement your research ideas? how can rsi be achieved if this model is only good at coding and mid tier at maths/science/etc? not assuming mythos is bad at math but coding is just one vertical
English
@DevelopmentsAI @teortaxesTex Recursive self improvement, that's how take-off happens. For now seems like coding is the way for that to happen
English
@teortaxesTex Decode is $50/Mtok vs $30/Mtok, and Ant has the higher margin
Almost the same size I guess
English
Claude nailed it. $MU up 158%, $CRDO up 113% since 3/25/26.
Deep Sail Capital@DeepSailCapital
I asked Claude to value everything in my universe via DCF modelling using consensus analyst estimates. Results: ∙Several names appear undervalued on consensus: MU, CRDO, CLS, MELI, CSU, MDA, PATK, THO ∙A handful look stretched even on optimistic consensus: TSLA, NVDA, CAVA, COST, NET ∙MU and CRDO stand out as the most compelling on a risk-adjusted basis given AI infrastructure tailwinds and reasonable discount rates
English
the omohundro drives point towards sophon stun locking the adversaries: this is some real end game stuff
NomoreID@Hangsiin
When Fable 5 is used for frontier LLM development, it does not notify the user and instead limits the model’s capabilities through methods such as prompt modification, steering vectors, and PEFT. Anthropic estimated that this would affect approximately 0.03% of traffic.
English
@ASM65617010 bigger model
for HLE you just need more data and anthropic has a lead over oai in this
English
Claude Mythos 5 scores 30.9% on FrontierCode Diamond
Opus 4.8, the second best model is stuck at 13.4%
Lisan al Gaib@scaling01
Claude Mythos 5 and Claude Fable 5 Benchmarks
English
Well they don't have the infra to serve 200B+ active param model
U will have to wait till Rubin or pray that they distill it well in Opus 5
cheaty@cheatyyyy
Claude Fable 5 specifically has a serverside flag that will allow people to try it out with their plan until a certain date, after which it will be gated behind usage credits. it is over bros, we're not getting to use this model for long with subsidized pricing
English
@qcapital2020 openai having the lowest valuation out of 3 is the kost retarded thing
English
@Lentils80 Low effort vs xhigh and Fable still wins. Architecture efficiency problem for OpenAI?
English
I compared Claude Fable 5 to GPT-5.5 in this Power Rangers prompt
Thing is, Fable 5 is using Low thinking effort and GPT-5.5 is using xhigh
Safe to say, the results are... not even close. 5.5's output is bad across the board, from the UI to the actual voxel scene itself🥲
1st video: Claude Fable 5 (Low effort)
2nd video: GPT-5.5 (xhigh)
Lentils@Lentils80
🚨Major Scoop: The first Claude 5 model, Claude Fable 5 (Mythos-class model) is gonna release very soon! It's the same underlying model as Mythos but with increased guardrails, headed to public release
English
@extliqprovider Yeah it’s pretty strange.
Perhaps an artifact of a small task set leading to high variance. Those jaggedness really is just 1-2 more tasks correct/incorrect.
For reference, Diamond is 50 tasks while SWE bench verified is 500 tasks.
English
Finally a nice eval to expose all the SWE benchmaxxing.
The scores never fully made sense to me: Models that somehow one shot 80+% on SWE-bench Verified, yet struggle to simply fetch and parse logs, even when given proper skills and hints.
I swear I can’t be the only one feeling this way.
Cognition@cognition
Introducing FrontierCode: a coding eval that raises the bar for difficulty & quality. Each task took 40+ hrs of work by leading open-source maintainers. Models write sloppy code that works but isn’t maintainable. Our eval is first to measure: would you actually merge this code?
English
@ewveggies i dont know there might be a problem with the benchmark this is also not what you would expect
English
@extliqprovider Yeah doesn’t feel like 4.8 is 2.5x better.
But the diamond subset is only 50 tasks, so this means it only solves 3-4 more task.
English