the eras of abuse https://x.com/stevekrouse/status/2092349508885549211/photo/1
Notes
- @whoisnnamdi Your post was like "this doesn't make any sense" and you were right! I'm a big fan of tokens becoming commodities I'm almost dreading a big new expense frontier model drop, which is kinda surprising. You'd think I'd be eagerly anticipating it I think it'd be a great time for…
Text may be partial · Read original
- @whoisnnamdi - have you written an update to your post about token economics? I think the shift to cheaper OSS has certainly begun!
- @chanjason Honered to be included
- @wafer_ai @s_pandey101 @vercel My agent confirmed: no cache hits with wafer. Would love tips there!
- @wafer_ai @s_pandey101 @vercel Then I did another demo with multi-turn where it was the same (or slower) than fireworks at 128 TPS I sent my agent to try to investigate what's going on... @wafer_ai - does wafer not auto cache? Would it be helpful if we sent over a minimal repro of the speeds we're seeing?… https://x.com/stevekrouse/status/2090217427938238601/video/1
- @wafer_ai @s_pandey101 @vercel I am definitely not seeing 1000 tok/s, and would love help if it's a me problem First I have a video of fireworks vs wafer doing a single tool call where wafer is faster but only 300 tps https://x.com/stevekrouse/status/2090216786738159946/video/1
- @wafer_ai @s_pandey101 @vercel woah holy crap i gotta try this
- RT @daniel_levine: wow... @juberti @paulcarletonjr @tobi @simonw ... @ValDotTown is your favorite developer's favorite developer tool Che…
Reposted by Steve
Text may be partial · Read original
- @originell @ellie_huxtable welp it stalled halfway through the first run, but running again!
- @paulcarletonjr https://blog.val.town/mcp-core-maintainers
- Pumped about how @paulcarletonjr & the MCP core team at Anthropic use Val Town! https://x.com/stevekrouse/status/2090172943955263681/photo/1
- @ellie_huxtable ok trying! i already pay for CleanMyMac, so wonder what else this finds
- RT @hipotluck: We were surprised to find FullEnrich at the 'pareto frontier' of email -> LinkedIn enrichment 40% match @ 5.8 cents per mat…
Reposted by Steve
Text may be partial · Read original
- @adamgordonbell i'm a big proponent of deepseek v4 flash and glm 5.2 i bet the newer deepseek pro and glm 5.3 are pretty great too but haven't played with them myself yet deepseek flash is crazy cheap and very smart
- @adamgordonbell i think people think fable is the frontier model, and i think that's mostly true i don't think there's non-frontier model people think is frontier it's more the other way around where lots of models are almost-frontier-quality but people don't yet know about them
- @dvsj_in ha! thanks!
- @dvsj_in again, kinda like with humans i only take the time to 'get to know them' when i 'hear good, relevant things about them' for example, i haven't spent much time with kimi k3 though i hear it's great - just not super relevant for me (townie) but yeah outside of townie, i'd be…
Text may be partial · Read original
- @CharlieMolthrop i like this idea, depends how you define wisdom one definition i like is 'ability to learn from other's experience' but i'm not sure that's exactly right
- @dvsj_in mostly vibes these days, honestly kinda like how i get a vibe for a human person
- @patife entertainment is hard
- i feel like i am so plugged into the pareto frontier someone ask me a question i gotta put this knowledge to use (before it's out of date)
- lol this was actually me reporting a bug! (now fixed) 4/5 = opus, sol, k3 5/5 intelligence = fable https://x.com/stevekrouse/status/2089805580835721484 https://x.com/stevekrouse/status/2090088484543353305/video/1
- @majamediaco @isabelunraveled So good!! Added to https://legibility.fyi/references
- @HodaNoorian It was a bug???
- RT @chrislakin: if you're going to distract, scrolling twitter is on the pareto frontier of distractions imo. everything else is either les…
Reposted by Steve
Text may be partial · Read original
- @simonlast Very cool that you're releasing it as an oss thing-in-progress I feel like notion would be particularly well suited for gbrain style RAG. Tbh I'm a bit surprised you guys don't already have that built in! I guess that goes to show that AI + notion tools is pretty good at picking…
Text may be partial · Read original
- "That's a real gap."
- @utotranslucence thanks! yes our first! she is due dec 24!
- @utotranslucence yeah, that's what we've been trying to so far and it's been promising! i feel like for you, i'd be curious to see a "job posting" sorta thing. i've been collecting my favorite most legible such "asks" here: https://legibility.fyi/ask ironically the very first entry is for an au…
Text may be partial · Read original
- @utotranslucence Did it work for you?
- @anderspitman why do you use it? why do the locallama folks use it?
- very surprised to see notion show kimi k3 as more intelligent than sol and opus 5! https://x.com/stevekrouse/status/2089805580835721484/video/1
- Breaking: the word "breaking" is being abused on Twitter for non-breaking news
- Folks better be careful with "breaking" on twitter; otherwise we gonna have to mute it
- @JakeZegil this is the correct answer faster models = flow state
- @JakeZegil to what end?
- cool to see qwen 27B taking over the timeline imo super smart small models are cool because: 1. feels like we 'own' it more when it runs on hardware we already own, or can afford 2. exciting to see the compressibility of intelligence! how compressible can we make it??? i was surprised to find that a dense model like this isn't cheaper + faster in the cloud though! MoE models like deepseek flash – 284B params; 13B active – run faster/cheaper on premium GPUs than a 27B dense model so in practice, i'm not really sure who qwen 27B is for. i don't expect to use it personally or know anyone who will, but it's really exciting to see frontier intelligence continue to get more and more own-able and compressed
- I'm surprised folks are not getting pwnd more often Given all the YOLO mode + MCP connections + vibe coding, you'd think there'd be a lot of new insecure software out there, ripe for script kiddies armed with openclaw bots finding exploits
- @utotranslucence I was thinking of posting something similar for a night nurse in NYC...
Text may be partial · Read original



