ทวีตที่ปักหมุด
Nik Cubrilovic
1.2K posts
Nik Cubrilovic
@dir
data + ai engineer. @squirrelscan_ @opennem. at https://t.co/yOGtW7uHIA and https://t.co/kvkcg5TrG9
Australia เข้าร่วม Mart 2007
534 กำลังติดตาม340 ผู้ติดตาม
Codex bros, please add passkey support to the built-in browser! A bunch of sites are completely fucked if you can't use passkeys (like this hellsite), and don't gracefully fall back to password authn, you're just fucked.
English
English
English
English
English
luxury ICE car models are going to hold or grow their value increasingly well as the economy adjusts to AI
exists in the same genre of things like land, rare wines, collectibles, etc — near-fixed supply in an ever richer and richer world
merp@0xMerp
very high conviction that these typs of cars are extremely good investments over the timeframe of decades most people think of them as toys but the world is changing, they literally just dont make cars like this anymore people pay a lot of money for time capsules
English
English
English
@jeffnolan @Electrarythm @sporadica also last one where you're driving the car, rather than driving the computer thats driving the car.
English
@Electrarythm @dir @sporadica The last naturally aspirated V8 from Ferrari. It is no longer undervalued. 2 years ago you could pick one up for $140k, now they are pushing $300k. 458 Speciale, the highest output variant, are selling $800k+.
English
@amix3k linting should catch that - best output on agents are when you have a tight feedback loop for it to call (format + lint + typecheck)
English
I used Opus 4.8 today and saw very little difference from 4.7. It still produces a lot of slop.
One example from today: it used inline imports in Python. These are generally a bad practice, but especially when they’re unnecessary, and our guidelines explicitly say not to use them.
Amir Salihefendić@amix3k
It’s ridiculous how much better Codex + GPT-5.5 xhigh is than Claude Code + Opus 4.7 xhigh right now. Claude Code feels like slop. I have to be super precise and careful just to get decent results.
English
It'll carve out a large part of the lower / mid market - but for F1000 it's either ant or oai because they're setup for enterprise (none of the Chinese model providers are in the same way - not that they'd host with them anyway). Same way there were always more affordable options to Oracle, SAP, Exchange etc., yet ..
English
The one thing which really will make a difference won't be this. It will be intelligence built into the apps (which has to call a model).
However, that will certainly cost the cheapest good enough model companies can find, due to the impressive costs it generates.
Models are commodities to a large extent. Once the initial euphoria passes, you use the cheapest that gets the job done reliably.
English
What's Athropic's coding TAM? The company's at ~$50b ARR, and my understanding is that most of their revenue is from developers that spend ~$5k/yr on avg.
Given there are ~20m professional developers globally, does this mean's Anthropic's revenue only has 2x-3x upside from here?
English
@ThinkFinance999 @lpolovets oh and security - audit tools burn more tokens for the code written than what it took to write the code. same with monitoring. code token burn is going to look quaint
English
@ThinkFinance999 @lpolovets code is probably still the #1 category for tokens used, but long term it likely won't be top 5 after sales/revops, support, compliance, marketing + content, etc. especially with user and transaction based agents now being deployed. so to address OP the TAM is .. large
English
in their series G announcement they attribute $2.5B of $14B to claude code:
anthropic.com/news/anthropic…
in their series H announcement 2 days ago they're at $47b:
anthropic.com/news/series-h
but for the first time no longer split out cc - likely because it's flattened out in growth terms vs enterprise / API (there are many other signals corroborating this - including my own experience across ant enterprise customers)
English
@JerimiahLee you get there eventually with improved conditioning - in the meantime you can slow down / walk / replace with stationary bike time.
English
@lpolovets ie. that dev headcount reduction will all go to tokens - either oai or ant (this is happening)
English
@lpolovets a) coding as proportion of rev. is at ~10-15% and shrinking, and will continue to do so
b) number of developers will shrink but total spend per developer will be *a lot* higher (already is!)
C) main top-line growth driver is token spend in enterprise non-coding roles
English
ok, here are the Opus 4.8 results for ErrataBench, a benchmark that measures how well LLMs can find and fix errors in English text. Some highlights:
1. Proofreading Quality - no improvement over Opus 4.7 and still lagging behind GPT. Performs the best with no reasoning.
English
@BenjaminDEKR very narrow definition considering large kabooms were common in Huntsville during development
English
It is actually kind of amazing that no Saturn V ever exploded on the pad.
Zero catastrophic failures, all the way through Apollo.
And they designed that thing with slide rules and string.
English