@bendersej @mattpocockuk Depending on the model you get between 70k to 150k of reliable instruction following.
English
TG
1.2K posts
@ossaijaD @mattpocockuk Not sure what you mean by « smart zone »?
I’m at around 33k tokens in context before the first message on a fresh session.
It’s not lean but it works perfectly for me, so I haven’t optimized it yet (for example I don’t use rules but layered Claude.md’s)
English
What 'advanced' AI coding techniques are you using?
I.e. what do you feel like you've discovered that no-one else knows about yet?
English
@mattpocockuk Not sure this is advance but session are best treated as ephemeral objects, my workflow is designed into phases each phase produces an artifact(.md) for the next phase.
That way I don’t get to fight context rot, context anxiety or whatever name it goes by this days.
English
@bendersej @mattpocockuk How does the context window smart zone not get filled up fast?
English
@mattpocockuk Thanks! Custom CLIs and MCP servers + Custom Telegram integration (for HITL)
For example one helper prints out the entire local DB schema + connection string by + ports of the services - checkout aware (I run multiple sessions in parallel in checkouts, not work trees)
English
@songjunkr I find working with Kimi K2.5 on @opencode zen better and faster experience than codex and opus 6.6
English
3개월 뒤에는 Mythos 수준의 오픈소스 모델을 사용할 수 있을거라고 확신해요.
지금도 사람들이 잘 모르지만, 오픈소스 모델로 Opus-4.6과 비슷한 수준의 코딩이 가능하거든요.
못믿으시겠다면 GLM-5.1을 클라우드로 한번 사용해보세요.
한국어
@theo It mostly likely all anthropic models based of Mythos like 4.7 will keep having this problem
TG@ossaijaD
Mythos Aggressivness if not stamped out will make it a very bad model to deploy for long horizon tasks
English
@badlogicgames Really spicy that one
TG@ossaijaD
Dwarkesh:Jensen is the first time someone has seriously challenged Ant worldview on AI, the latest been if China was first to Mythos like model they will use it against the world but we will us it responsibly so we should be the only one who get to make and control them.
English
Many such cases
Brendan Falk@BrendanFalk
The fraudulent job candidates aren't even trying any more
English
If AI is indeed a 5 layered cake why should winning only in one layer best for the US?
TG@ossaijaD
Dwarkesh:Jensen is the first time someone has seriously challenged Ant worldview on AI, the latest been if China was first to Mythos like model they will use it against the world but we will us it responsibly so we should be the only one who get to make and control them.
English
@XP_Ehsaan @Real_SilverLine @songjunkr @AnthropicAI As for standing your own inference stack I hear this is not as simple as it seems unless you have the engineering resource to pull it off. You may however want to consider other providers zen from opencode is great, gives you access to many of the latest open source models
English
@XP_Ehsaan @Real_SilverLine @songjunkr @AnthropicAI At this stage of model development being able to give yourself optionality is key. Going in on opus 4.6 isn’t int. I think since this year other models have performed better than anything that has come out of Ant so far. Remain to be seen where 4.7 lands eventually
English
Opus 4.7 토큰 테스트
토크나이저 차이로 제미나이의 2배를 사용합니다.
Opus 4.6 대비해서도 50% 많이 사용해요.
이건 사실상 같은 한도에서 모델이 50% 더 비싸진겁니다.
한국어
TG retweetledi
Shocking result on my pelican benchmark this morning, I got a better pelican from a 21GB local Qwen3.6-35B-A3B running on my laptop than I did from the new Opus 4.7!
Qwen on the left, Opus on the right
English
your code is part of the prompt btw
if your code sucks your outputs will continue to suck
you have to improve the code otherwise you will never escape the gravity well of all your bad patterns and tech debt
more true with autoregressive LLMs than with humans
English