Shane

[Yun-Ta Tsai @yunta_tsai]

I made a video about what happens when you execute `c = a @ b` in PyTorch, from the code to silicon.

Hopefully it will help you appreciate that simple line of code more.

Made by Grok @bot. Enjoy!

[Shane @BehindM37983]

@_can1357 @chasen_liao omp2 什么时候发布啊🫣

[Can Bölük @_can1357]

Dario only seems to care about the IPO price, Sam wants to have the best/fastest/cheapest models. Both of these companies have one real product: the model.

OpenAI will proudly admit this. Anthropic, on the other hand, tries to control the products around their model as well (promising monopoly on all industries comes with more 0s in their future valuation).

They are absolutely horrible at making any product whatsoever, but ofc this doesn't matter if:
- a) the model makes it a unique experience they cannot find elsewhere for the user
- b) they bully competition out of the market by offering API at 25x the price of a sub, and not letting people use their subs freely: no sane person will use cursor if it costs 4k$ a month, if claude-code does it for 200$

= Cursor heavily relied on Sonnet 3.5 for their product; Anthropic released Claude Code as a response a few months later.
= Figma, Claude Design (anybody remember that? no right? yeah...)
= They don't want 3rd party harnesses to exist, and tried to prevent their use in their subscription offerings quite a few times. [OpenAI came out in support and offered it freely instead]

Same deal with safeguards:
= cyber classifiers -> they have security offerings
= bio classifiers -> they mentioned getting into bio & Dario even has a bio bg IIRC (see latest drug development thing they published)

Funniest one is the ML research one, where they literally admitted silently downgrading the model if you were using them to train other models as if it was some legitimate US-China thing; no, they just don't want competition.

[Shane @BehindM37983]

@cenfangyu17 @thsottiaux same with you

[Shane @BehindM37983]

@thsottiaux So now on the web, are you also changing the model in the background without the user's knowledge?

[Shane @BehindM37983]

Pro 20X kept swapping my 5.6 Pro to 5.5 mini without telling me. Now on Plus, I select 5.6 High — no reasoning shown, instant answers. Asked the model: '5.5 mini.'

So now even regular 5.6 gets silently downgraded?

I paid for 5.6. If you can't deliver it, be transparent！

[Jun Song @jun_song]

Tens of thousands of users, including myself, are getting hit with massive rate limit nerfs even when only using Codex.

So their official stance is basically just treating users like idiots?

Classic response from a company that does nothing but chase Anthropic's tail.

[Tom Huang @tuturetom]

我们把 DeepSeek V4 Pro 和 Flash 的使用门槛，直接打到了首月 5 美元无限用🚀🔥

正式推出OpenDesign Go套餐！这可能是目前使用 DeepSeek V4 最便宜的方式，且不限 Token，不额外收费。

首批内置 8 个热门模型，包括 DeepSeek V4 Pro、DeepSeek V4 Flash、GLM 5.2等模型。套餐额度每 5 小时刷新：每 5 小时约 16 美元、每周 40 美元、每月 80 美元。同样是首月 5 美元，OpenDesign Go 的使用额度比 OpenCode Go 高 33%🔥

过去一周，在经历DeepSeek大涨价的情况下，我们依然开启了为期两周的DeepSeek V4免费无限用的活动。短短一周大家就在OD内用掉了价值约 1.8 万美元的token额度，这种热情也完全点燃了我们，这次我们把原本为期两周的无限活动，直接变成 Go 的长期权益💥

我们的目标是让 DeepSeek 真正变成 CheapSeek，让更多人可以用最低的门槛，通过OpenDesign做出顶尖��计作品💪💪💪

[Xiaoyin Qu @quxiaoyin]

Dear Dario,
1. If Claude can cure cancer to save people like your dad, why should we "pace the progress"? Does that mean more people with Hepatitis C will die?
2. If Fable is so cyber-capable that it must be restricted, why are its safeguards too dumb to distinguish cyber defense from cyber offense prompts? When Hugging Face was under attack, why did Fable refuse to help the defenders?
3. We’re glad you want AI to cure cancer. Why is it OK for Claude to force 30-day data retention on pharma's own data and start competing firms but NOT OK (IP theft) if others distill Claude's data?
4. You’ve said advanced models can recognize when they’re being tested and change their behavior accordingly. So why is government(or anyone) able to design the most thorough test before every model launch? Would that just encourage manipulative models?
It seems that every one of your “safety” proposals seems to end the same way: Anthropic gets more leverage, ordinary users get less access, customers pay higher costs, and competitors bear higher regulatory costs, maybe people are not distrusting AI, they are distrusting your approaches with AI.

[Shane @BehindM37983]

@BlockedPaths I don't agree. I think it's a good open source harness compared with other ready-to-use ones

[Matt Pocock @mattpocockuk]

The outrageous effectiveness of Leitwörter

I've realised that all of the great skills I've written share one thing in common.

They make heavy use of Leitwörter - leading words.

A leitwort comes from literary theory. It's a repeated word or phrase used throughout a text to establish a theme or anchor meaning.

In skills, a leitwort is a word or phrase the agent uses to guide its own behavior. In other words, it's a word that leads the agent in a certain direction.

Let's take the leitwort "zone of proximal development" from my /teach skill. It's a phrase from the study of education. It means the "zone where the user feels challenged but not overwhelmed".

I use this only a couple of times throughout the skill's SKILL.md, but I've seen it almost every time the agent invokes the skill.

- "Let me adjust the lesson so it's in the user's zone of proximal development."
- "I'll read the learning records to establish the user's zone of proximal development."

In other words, that single phrase encodes how the agent should behave, in a concise token the agent can itself repeat to reinforce its own behavior.

Not only that, but it also likely tickles the agents' parameters related to educational research and "being a good teacher".

For engineering, leading words like "tracer bullets", "deep modules", "test seams", "clean code" are outrageously effective for leading the agent to produce better code.

So a leitwort in AI is any word or phrase you use that appears in the agents' thinking traces and guides its behavior.

Enjoy finding your own.

[Shane @BehindM37983]

@thsottiaux Copy markdown formula and table with one click