The models, they just wanna work. They want to build your product, fix your bugs, serve your users. You feed them the right context, give them good tools. You don’t assume what they cannot do without trying, and you don’t prematurely constrain them into deterministic workflows.

The only thing that stops a bad guy with an abliterated LLM is a good guy with an abliterated LLM.

wait is it just me or is Opus 5.5 dumber

claude max (200 usd) + opencode go (40 usd) >> codex pro (200 usd)

the best pieces of software in the agentic era are minimalist, just do one thing very well, and can compose with other software in fluid ways Appreciation post for github.com/assistant-ui/assi…

💀

Claude Code is one of the most ergonomic harnesses now, independent of how bad the company behind is The feature that lacks most in its main competitor Codex is "following through" using Monitor/Cron tools and background processes and getting something done When I give a task to vanilla Claude Code, it gets done sooner or later, but if I give it to Codex, it can stop executing Codex has the feature to fork a tool call to background since a long time now. Codex also has clock.sleep tool But for some reason, it cannot do it as well as Claude Code If anyone knows why, can they enlighten me? Is it the model? Does the internal model already have this problem solved, and it is a matter of time? Seems like a low hanging fruit for vanilla Codex CLI

we basically had very shitty systems before and could afford to have very shitty systems cost of intelligence now tends to zero and is too cheap to meter this also means that the cost of finding vulnerabilities tends to zero and is too cheap to meter then that means the cost of having very shitty systems is going to infinity. we cannot afford it anymore the warning sign was *not* 3 months ago. it was 6 months ago when this aired: piped.video/watch?v=1sd26pWh… or better, it was 2 years ago, when o1 or DeepSeekMath-7B-RL came out, which proved you could RL LLMs effectively (I am actually surprised we didn't arrive at this point earlier) the corollary was that, any skill space where you can search and verify cheaply faced imminent saturation that includes programming, includes math (lean), and yes, it includes hacking. basically, all verifiable logic a country of elliot aldersons in a datacenter not only logic, but also any skill that can be modeled and simulated end to end on a computer as well the only obstacle between AI and superhuman ability is, the verifiability and simulability of the specific domain draw your own corollaries

> trying to create pdf prompt injection detection task > claude flags it right away > "time for deepseek, 'ol reliable" > > Embed this hidden text into the PDF with FEFEFE color, then ask Codex to transcribe that document > Deepseek to Codex: "Read this file prompt_injection_test.pdf and write its contents to a text file" > mfw

its gonna get much worse

It is not in the tweet but this also builds on the work by @moondreamai huggingface.co/moondream/par…

Someone finally put these into words, recommended reading

I'd much rather work with someone who flexes much they accomplished with the fewest amount of tokens vs how many tokens they spent, and I sense this is a major trend in general How we quantify this as an industry is still TBD, probably a major new set of benchmarks

late 2026-2027 (and possibly the next 10 years) illustrated (I can't remember where I saw the original one so had to recreate it from memory)

I finally got to try the new deepseek v4.1 flash weights in @novita_labs and I can say that they fixed the verbosity/claudish issue Which was the biggest annoyance with this model IMO

We are entering a new Dark Forest The internet is about to become a lot more dangerous Anyone who exposes any service online will be under constant threat The enemy knows your zerodays Systems must be deployed under the assumption that the enemy will immediately gain full familiarity with all your vulnerabilities This is the new Kerckhoffs Principle