@1a1n1d1y Has happened to me a few times, but never anything destructive
Niklas Sheth@niklassheth
Opus 4.6 is so proactive that it answered its own question for me
English
Niklas Sheth
1.1K posts
okay here is my comment
very very often after 4.6, opus will ask a question or make a suggestion, and then just do it. this extends to a lot of things it does, but electing to destroy instances costs me actual real money and it's getting super tireseome, so i'm shaming publicly
English
file indexing, code indenting, and basic file operations are broken beyond comprehension, so i'm back to vs code
wukko@uwukko
@cheatyyyy yes, and i just switched to it again, because my macbook would burn otherwise
English
I rigged up a low budget OpenClaw with the Claude Code Discord integration, works almost as well and stays within the subscription's terms
Boris Cherny@bcherny
Starting tomorrow at 12pm PT, Claude subscriptions will no longer cover usage on third-party tools like OpenClaw. You can still use these tools with your Claude login via extra usage bundles (now available at a discount), or with a Claude API key.
English
@natolambert I wonder how long it will be until frontier models are no longer available through a text completion API. I'd guess 2 years
English
This was actually already policy.
Regardless, destroying demand was coming with undercapacity and increasing verticalization/integration is the right move.
Perfect move in fact, despite people being understandably mad.
Boris Cherny@bcherny
Starting tomorrow at 12pm PT, Claude subscriptions will no longer cover usage on third-party tools like OpenClaw. You can still use these tools with your Claude login via extra usage bundles (now available at a discount), or with a Claude API key.
English
@jaxgriot @leothecurious It's impressive for its size, it did well on basic computations like solving a quadratic equation. Knowledge is severely lacking of course.
English
is this the most "overtrained" LLM out there or what? if this really works and doesn't hallucinate as much as i expect it to at this scale, there are so many practical applications where this thing can be put to productive work at unmatched efficiency.
Liquid AI@liquidai
Trained on 28T tokens with scaled RL, LFM2.5-350M is a step change from LFM2-350M: > instruction following: 18.20 → 40.69 > data extraction: 11.67 → 32.45 > tool use: 22.95 → 44.11 These are the capabilities that matter in production.
English
The Claude app and Claude Desktop are so janky that I switched back to the CLI. I appreciate how fast they’re shipping but it’s not a good look
English
@casper_hansen_ What batch size is the 300 tok/s at? If it's 1 then that's not very economical for 8xB200, about 1M tok in an hour with 8xB200 @ $5/GPU/hr = $40/MTok
English
every inference engine should have a section in their docs with exact commands to achieve best possible tokens/s on the most popular models
i'm told kimi k2.5 can run at 300 tokens/s on B200s if you run nvfp4 with speculative decoding in open-source
English
@JesseTayRiver Surprising that they'd document it instead of labeling them as "for manufacturer use only" or something
English
Air conditioner and heat pump manufacturers often reuse the same unit across multiple capacities and put limits on the control board. Want to switch your 2 ton air conditioner to 3 tons? Easy as flipping a switch
English
@skydotcs I can tell that's 5.3 Instant, use GPT-5.4 at least
English
@cheatyyyy @krapstarr @jeremyakahn @FortuneMagazine @beafreyanolan Because that was a step change, my point is that before this leak they could've released an Opus 4.7
English
@niklassheth @krapstarr @jeremyakahn @FortuneMagazine @beafreyanolan why completely leave out 4.1 to 4.5 lol
English
Exclusive: Anthropic left details of an unreleased model, exclusive CEO retreat, sitting in an unsecured data trove in a significant security lapse. Great reporting from @FortuneMagazine's @beafreyanolan fortune.com/2026/03/26/ant…
English
@jrysana @daniel_mac8 Yeah you can serve Deepseek-V3 at around that speed per GPU
English
@daniel_mac8 In other words, *per-GPU* this represents a total of ~11.5k tok/sec across e.g. 230 users each getting 50 tok/sec. Which is good but not atypical or a "warp speed", by which I assume you believed it to be a substantial leap over the status quo.
English
a gcp eng networked 96 b200s and ran qwen 3.5 27B at warp speed of 1 million tokens per second
did you know you can do that?
English
@krapstarr @jeremyakahn @FortuneMagazine @beafreyanolan No? 4 to 4.1 was incremental, 4.5 to 4.6 was incremental. There was no reason to believe it would be a step change
English
@jeremyakahn @FortuneMagazine @beafreyanolan TL;DR - the next model will be a step change which we all already knew
English
ARC-AGI-3 is basically impossible for LLMs right now because there's no inter-frame compression, so it consumes context really fast. Tokenization is so engrained in LLM development because it's simple and effective, but we're hitting the limits now. New ideas are needed
English
@Setuna7777_2 That is odd, especially since Muon pretrained models (like Kimi K2.5) do better with Muon as the SFT optimizer
English
Is there a reason Muon wasn’t used in Composer 2 training?
Cursor@cursor_ai
We're releasing a technical report describing how Composer 2 was trained.
English
GPU market is pretty dire these days, huh?
English
@DrugGovoruna @3RenChengHu @moultano Yes, and a calculator was provided for the quantitative section. I guess most people don't know how to convert between square units
English
English
I wish this was contextualized with something like "At each 10%ile of reading ability, here's a test question we'd expect half of respondents to get wrong." Apparently average reading ability is 7th-8th grade, but more than half of of adults have attended college?
English