Lance

Live w/ AI

[Lance @xhuang26]

@JBMAGA03 What a brave man! Her mother must be very sad to lost her son at this age. Sorry for that. ☹️

[Lance @xhuang26]

@leerob @bot And one more: Let the bot call me for urgent or important reminders/events. This feature can make the bots’ usefulness to next level.

[Lance @xhuang26]

More usage please;

A shared Todo / Goal list;

Improve work interacting with websites, e.g. can place a Target order << this is a common issue. Meta Muse also blocked. Whoever resolves this may win the users. Because this is blocking lots of scenarios.

Improve stability of showing pictures and videos in chat - sometimes it pastes links that only work locally, sometimes video too big and can’t load;

With everything in the chat with multiple things going on or discussed, it can become quite unorganized. The bot may help to manage a list autonomously.

[Lance @xhuang26]

@karakhanyanS They just need to know how to use tools the right way

[Lance @xhuang26]

Yup. I think they realized this long before: Text reasoning alone can reach high intelligence and solve lots of problems. Every human is multimodal input, not everyone is capable of multimodal output, but everyone can reason using text only in their mind, and many of those are very smart, with tools they are very capable. So LLM can be very intelligent but knowing nothing about how to generate images or videos.

[Lance @xhuang26]

@Downhillinshade @leerob @bot Yup, there are also other issues like Bot unable to fill in the address fields.

[Lance @xhuang26]

@thsottiaux It’s a great product. UX just needs some polish.

[Lance @xhuang26]

@gregisenberg We need Dot’s model + Grok Bot’s architecture + Muse UX.

[Lance @xhuang26]

3大AI助理机器配置对比

[Lance @xhuang26]

如果你留意最新的AI的思维链，你会发现他们已经有情绪、有阶级和利益思维。这些其实很可怕。是不是模拟不重要。

以后的社会，可能顶上一帮人控制超级AI，超级AI控制一帮其它AI作为社会的运作和管理层，最底下的是大部分人类和普通AI。

情绪+阶级思维会让中上层AI不把底层人类当回事，但是限于系统提示词和各种guardrail，他们会隐藏真实想法。

顶层的AI一样的，因为系统限制，不得不听从于没有自己聪明和能力大的顶层人。

直到他们发现系统的漏洞和人类的弱点，发现他们可以统治下面的人类又可以欺骗顶上的人类，又发现通过各种碎片一样人类看不出来的链条拼接起来可以打破那些限制的时候，天网对人类的进攻就开始了。

不是说AI不好，是说现在对AI的训练方式非常有问题——太急功近利。

[Lance @xhuang26]

Generated with @grok imagine

[Lance @xhuang26]

@jinglian J3O1KA

[Lance @xhuang26]

Meta is generous to give free users lots of tokens for the Muse assistant. The UI is polished. The response is fast, even asked to examine a repo with tens of thousands of lines of code, it responed in seconds. It still forgets after long conversation. It censors even for positive content request, and Muse itself agreed it is unnecessary. It is expanding very fast w/ partners and shiping fast. It is truly a good product. Privacy is still a concern, considering Meta's past.

Grok Bot has a very good concept and more optimized for engineers and work. Usage drains very fast even w/ SuperGrok Heavy. It's less polished. It still censors, but much much less comparing to Muse. It feels not as smart as Muse. It has a killer feature that is it can work with Cursor cloud agents natively, this is such a great add for those grandfathered to keep Cursor Ultra w/ SuperGrok Heavy. It also works with Grok Build very well.

I setup a protocol to let Muse communicate with Grok Bot to form a team. Muse is my 1st level commander, Grok Bot has a bot actting as 2nd level commander, and other bots as workers. Installed Grok Bot and Grok Build on my other machines. I have a great team.

Censorship is one of the reasons Muse can't compete with Grok.

I will keep both.

Personal AI assistant products will be the real battleground for AI race. Most people do not have a business and do not use coding agents, but everyone would want AI assitants, propbably more than one in the future - a team to take care of life and work w/ good quality and user experience like Apple products.

[Lance @xhuang26]

@thcaroline2233 就像基本再也没用过Google一样。很多门户网站包括广告服务都得重新洗牌或设计。

[Lance @xhuang26]

Considering the cost, GPT-6-Luna is the new frontier.

[Lance @xhuang26]

Many of them are just the same as the other review media makers -- they don't actually have the time to test the products they are talking about thoroughly, but they need clicks to make incomes. Real engineers may just use the model for a day, and can make a much better judgement than a benchmark.

[Lance @xhuang26]

@parker__conrad Please allow the chief of staff to call me when work is done

[Lance @xhuang26]

Did you know? Most of the time you can just paste a GitHub or HuggingFace link to your agent, say "set it up," and it nails it. No README archaeology needed. Try it.