傅剑寒

做对且艰难的事，把艰难的事做对。
探索自己的边界

[傅剑寒 @FragileOffice]

@vista8 claudecode删减了80%系统提示词的确在沿着Bitter Lesson所说的方向前进，随着模型的增强，和它搭配的harness也要与时俱进。superpower最近也在删减一些限制条件来给llm松绑，这就是skill的未来趋势，skill会逐渐从强制指挥转向指引方向，而让llm跑马圈地，尽情释放它的性能。

[傅剑寒 @FragileOffice]

@CuiMao 下次能让牢梁召唤鲸鱼出厂吗，很期待通灵之术

[傅剑寒 @FragileOffice]

@kasong2048 对于通用大模型和通用agent来说，总会有一些是特定横切面的能力是���应该放在通用agent里的，比如grillme这种拷问能力，是一种独特的系统提示词一般的元能力，总会有些skill会被淘汰，但是元能力skill不会消失。其实我觉得grillme应该成为每个通用agent的系统skill来让大家使用，和plan是并列的提问方式

[傅剑寒 @FragileOffice]

@omarsar0 Your post leaves me with one question: when turning a task trace into a reusable skill, how do you distinguish the details that represent generalizable capability from the details that are merely noise or specific to that particular situation?

[傅剑寒 @FragileOffice]

@7qD1aFUQaw41485 推荐一下我做的增强grill体验的skill，外嵌一个map，让llm知道当前的步骤状态
主要解决三个痛点：
1 容易被大量问题淹没而抓不住重点和主线
2 当在一个话题的话题深扎时如果有灵感无法及时记录下
3 grill最后倾向于做mvp，如果不主动记录完整计划内容经常handoff中被磨损掉
https://t.co/S89Zvr5MGQ

[傅剑寒 @FragileOffice]

@mattpocockuk I hit an ambiguity using /to-spec on an LLM skill. It grouped 10 acceptance behaviours under an “Agent E2E” seam, and downstream agents treated them as permanent E2E candidates. This blurred acceptance scenarios, test seams, and maintained E2Es. Is that intended?

[傅剑寒 @FragileOffice]

@mattpocockuk Matt, my main issue with GrillMe is that the full plan gets lost when turning discussions into an MVP. After finishing it, I have to reconstruct the original inspiration from scratch. Grill to Does has the same problem. Will the new Grill fix this?

[傅剑寒 @FragileOffice]

@mattpocockuk I’ve been using teach and really like it. One thing I noticed: it’s good at making the next lesson, but not as good at knowing when to stop and switch to review or real practice.

[傅剑寒 @FragileOffice]

@AYi_AInotes 老哥做到了我想做的事，👍

[傅剑寒 @FragileOffice]

@LanLance24 @syhily grill-me是神级skills真的好用，遇到什么都可以用

[傅剑寒 @FragileOffice]

@yanliudreamer 我用过别人做的感觉最喜欢的一个是@mattpocockuk的grill-me。简单的对抗性提问->大道至简，凡是我有点模糊的时候就用。

[傅剑寒 @FragileOffice]

@addyosmani 这个我得拿出找资源的干劲去读了！

[Addy Osmani @addyosmani]

https://t.co/hIe0UX7z6T

[傅剑寒 @FragileOffice]

@yanliudreamer 男的也要露脸吗，我有点社恐。😟

[傅剑寒 @FragileOffice]

use your're post to study English!!!🥰🥰🥰@aleabitoreddit

[Serenity @aleabitoreddit]

I think my personal style of investing is a bit different, just some reflection:

It's inherently discretionary, based on stuff markets don't know yet. And a culmination of life experiences?

If you look at $AXTI, $RPI, $SIVE, $IQE and others.

Lot of it is guessing on unstructured relationships then seeing if it's right or not down the line.

$RPI is the perfect example:

1. Nobody really thought of Raspberry Pis for AI growth. Mainly people bought one or two just for class + education + hobbyist.
2. After OpenClaw, just noticed all my friends and people just buying Apple Mac Minis / RPIs for AI applications.
3. Found validation of that trend online with lot of people sharing video tutorials on AI orchestration with RPI.
4. AI was their ideal perfect growth vector, did some modeling, and thought it was compelling.

Earnings comes out and I was right.

Everyone in media was calling it a meme stock because there's nothing online that shows revenue growth from AI (was 14% forecasted revenue growth, turned out to be 58%, my projection was around 55%).

So it was a mix of guessing next industry trend (AI using lightweight hardware instead of GPU clusters), real life trends, then revenue forecasting off my guess.

For stuff like $AXTI:

1. Everyone called it a joke when I bought at ~$12. LLMs would hallucinate and say "hyperscalers/govs would have known about this by now and fixed this vulnerability with InP substrates"
2. Or would conflate very nuanced parts of InP substrate stack, where there's multiple different chokepoints in upstream processing.
3. So part of this was just discretionary based on what I've seen over InP substrate breakdowns, industry trends, etc.
4. Then also guessing the major supercycle was photonics (this was before everyone caught onto $LITE, and others). Or before you saw the $141B TAM projections from GS.
5. AXT owned 40% of InP supply chain, without them the supply chain just gets cripped).
6. All the "analysts" were forecasting steady InP substrate growth, few hundred million TAM, etc. or export controls.
7. Everyone kept trying to say $AXTI was overvalued based on TAM estimates. But if it's a few hundred million TAM you just think that's a joke and go into game theory over allocations.
8. Then I just had to guess, how much would this be worth if it were a NAND style bottleneck, what MC could it reach based on control, how much would hyperscalers price it as, etc.

A lot of the current research outputs from Goldman Sachs, or earnings reports from the Epiwafer companies, were confirmed after I published my piece on AXT. If you did research back then, lot of the same material /framing wouldn't have come up.

With stuff like $XFAB as you're seeing now, a lot of it is just pure guessing:

1. Not really any CPO materials, how much their MTP process makes in revenue, etc. Everyone online keeps saying they're not a photonics player.
2. But if you go through ASE docs or Gov websites, they all kinda cite XFAB as a major emerging player here.
3. $NVDA also evaluating them right now (maybe it's successful who knows).
4. No clear revenue around this area because their main silicon photonics process is still precommercial, but if you guess it's trying to create a EU supply chain to compete with $TSEM, once pre-commercial shifts to commercial, maybe similar but less volume contracts?
5. Then just seeing updates over the next few months to see if anything confirms this thesis guess.

_

I think a lot of information discovery still can be done with LLMs I'm seeing online. But it's also really hard to make a bunch of unstructured inferences based on unrelated material or even just trends you're seeing in real life.

So probably better to just do what's standard, eg. do valuation forecasting based on current numbers

Stuff like $AAOI, if they're projecting $471m/M h1 2027 and you see MC at $12B, probably undervalued might be a good idea to go long for next years.

Stuff like Samsung Electronics is easier, see what people are modeling for operating profits for 2027, 2028 then just seeing if it's undervalued or not at current levels.

Maybe something harder is $JBL. I haven't really seen any great volume numbers around 1.6T LRO, but you can just make a guess on how popular that might be then project how that might impact current MCs.

Or picking just good names everyone kinda agrees like $TSM, $INTC, $MRVL is also solid.

So a lot of things is just building up your life skills then applying that to markets. I don't think it's that can be taught with courses and stuff.

Of course, much of what I'm doing is just high conviction inference based on unconnected parts. Could always be wrong.

[傅剑寒 @FragileOffice]

@CuiMao Cuimiao：“我去过英语区，那里的AI coding很纯粹！”

[傅剑寒 @FragileOffice]

@AlchainHust 我装了好多skill，但是我除了一些我常用的，其他我根本不知道什么时候来调用，每个skill的上手成本真无法去衡量。

[傅剑寒 @FragileOffice]

@Khazix0918 就算两大强国都延缓ai发展的进程但是欠发达国家呢，为了实现经济增长，我觉得朝鲜都会举全国之力发展ai。并且想要倾向于发展ai的研究员会选择他们能发展的地方

[傅剑寒 @FragileOffice]

@CuiMao cuimiao老师能不能做一个Mr. Dario, we miss you.

[傅剑寒 @FragileOffice]

@CuiMao 霸气护夫