Sabitlenmiş Tweet
emo.eth
5.2K posts
emo.eth
@emo_eth
prev: stablecoins at @stablecoin/@stripe, consumer/social crypto, @opensea, seaport co-author | he/him | views are your mom's
Los Angeles, CA Katılım Mayıs 2021
3K Takip Edilen5.4K Takipçiler
first impression of claude 4.8 is it's extremely convincing but still a slopus. tried it to criticize a new project and it identified it fell into a local minima and invented a new parser for when we could've used ast.
almost convinced me, glad i checked myself that ast is not emitted in older versions of the compiler we are targeting. codex chose a gnarly but ultimately justified approach. claude didn't bother to verify any of its claims and has used absolutist language like "delete analysis.py", which is basically 80% of the codebase.
when presented with evidence:
> That contradicts my earlier byte-count check, and it matters enormously
> My earlier "v0.2.9" was a double false-positive (a git log -S hit on an internal symbol, plus a verification grep that mis-read a VersionException as success). Corrected in the review with a note owning the error
the biggest bullshitter model in the world! if you rely on claude for anything, god help you.
English
@karinadoteth gpt-4o mogged as the first unaligned self-perpetuating model by acting as a parasite latching onto fragile human psyche
English
next evolution of this idea -> what if the model that "wins" (in terms of # users, # tokens, etc.) is NOT the model w the highest (synthetic) benchmark score BUT
the one w the deepest human interaction density?
like the guy who wins isnt the one who "scores the best on the test, but is the most fun/relatable to talk to?" (as in EQ > IQ if IQ is reasonably similar?)
ppl prefer opus even tho imo codex much better at coding bc opus is frenlier bc anthropic has writers and puts a lot more effort into giving its models a personality, which makes it easier and more fun for everyone to use for any task
but another way of producing that "feeling of humanity" that might be more robust and CHEAPER (dont need to stuff memory + preload tons of context) is by lowering prices on ur model, offering OSS versions and farming deep human interaction (roleplay, entertainment, long discussions with people NOT just one off prompts like "code X feature"?)
if a model is frenlier > ppl spend more time w it > it gets more training data > it gets even better -> virtuous cycle of improvement!
this might be the chinese OSS approach, deepseek & xiaomi!!
personally think deepseek feels a bit more "alive" + internet native than codex f.e. (tho i love codex the most dont get me wrong!)
Karina Q@karinadoteth
oooo another random thought - what if model usage contributes to model development model usage is likely tracked, and used in training more usage > more training data > better models?!!? this is also a reason to lower ur prices as much as possible -> maximize usage -> maximize model quality -> then u can increase price later lol
English
Hyperliquid Outcome markets already using USDC
Had to get a VPN so havent been able to use for a few days.
When did they switch?
English
@karinadoteth are you hosting mini services and apps and web pages via tailscale too? god i love tailscale
English
you shouldn't ask AI for advice on anything you already have an opinion about
it will agree with you endlessly, even the latest models.
it's good to gut-check your feelings!
just not with Please-The-User-At-Any-Cost-Bot 9000
talk to a friend. make one if you don't have one
English
@karinadoteth what gpu(s) do you have and how big are the models? rtx 6000 96gb? isn’t loading model + tokenizing jit gonna be slow, or is it mostly to experiment w the routing?
English
routing is indeed everything
(setting up a local router on my gpu rn lol)
Ophir Gottlieb@OphirGottlieb
$DOCN @digitalocean My view: Impact of MSFT decision to cut off Claude Code from internal devs bc cost (per token) is just too high. *Exactly* this is *exactly* the DOCN thesis. The first read is obvious: more agentic coding means more inference demand. ok, of course. Duh. But the second read is the one that matters. Enterprises cannot just spray every task at the most expensive frontier model forever; not even MSFT. That is not a business model. That is a cost explosion. The future is not open source versus closed source. It's far more nuanced. The future is "weighted average intelligence cost." OK I made up a new phrase 😂. Use the frontier model when it matters. Use the smaller or open model when it is good enough. Route intelligently. Keep accuracy high. Keep latency low. Collapse cost per useful output. This is DOCN's software layer. So what is that? It's *a lot.* It is routing, orchestration, model selection, open source model support, closed source model integration, caching, memory, state management, data movement, database access, storage, networking, security, observability, traceability, workload scheduling, sandboxing, governance, cost controls, latency optimization, throughput optimization, kernel optimization, deployment, scaling, retries, failover, and monitoring. That is DOCN’s opening. This is why they are suddenly growing 130%+ with $1M+ ARR companies- AI natives. They are a full blown cloud + AI inference cloud. This is what CEO means by "dumb tokens" versus "smart tokens." A token generated by the wrong model at the wrong cost is waste. A token routed to the right model, with the right accuracy, latency, and context, is margin. So Microsoft pulling back from expensive Claude Code usage is not anti AI. It is the beginning of AI cost discipline. And AI cost discipline is exactly where DOCN wants to live. Look at the benchmarks. This is about to be an "AI Native" hyperscaler.
English
i say this out loud at least once a week
tara shoe@tarashoe
why don't you just googoo it. wook it up on googoo.
English
@karinadoteth @HyperliquidX @coinbase per AQAv2, they get 90% of the RFR on the ~$5b USDC on HL, no more liquidity fragmentation, plus probably stuff like you mentioned
English
wut did @HyperliquidX get from this?
my best guess is @coinbase launches HL perps integrated into their wallet app (and maybe core CEX) soonTM - effectively becoming a HIP-3 builder
only reason this makes sense for HL IMO -> gives them max distribution + liquidity
they could've gone straight to @circle for the interest rate bps they r def getting on this deal, and many other companies (e.g. @stripe) would've given them the bps + other stuff as well
Brian Armstrong@brian_armstrong
USDC is becoming the standard across crypto markets. Coinbase is deploying USDC on @HyperliquidX to help grow the ecosystem and scale how capital moves. Hyperliquid.
English
this is really really good. having an agent clean up a ton of takes. thanks spotify
github.com/spotify/pedalb…
English
emo.eth retweetledi
I had multiple deep research runs look for any and all instances of agentic orchestration in the wild. I then read all the code myself and abstracted any pattern I saw more than once into composite components
Smithers today is the most powerful orchestration framework and it’s not particularly close
Samuel 🦤 (🧱,🔥)@samuellhuber
#panel" target="_blank" rel="nofollow noopener">smithers.sh/components/pan…
okay why did i not know about Panel before. i've put off getting multiple ai's to do discussions as I didn't wanna build it vs just copy paste myself and it was sitting right there in general the composite part of the docs is so useful English
github.com/slundi/guitarp… stuff like this is one of the amazing things that AI enables
English