yantao006

独立开发者，前大厂技术专家 ｜现在一边学习，一边从 0 到 1 打造 AI 产品，探索更高效的 AI 工作流 ｜分享记录产品构建中的尝试、踩坑、思考 ｜ 新人博主，坐标深圳，欢迎交流

[Kun Chen @kunchenguid]

used grok 4.6 for a full day of real work, here's my unbiased review

i used it as my @myfirstmate - the orchestrator agent that i directly talk to that manages all my other agents - and also as a worker agent directly on some implementation tasks

first, the good:

yes it's fast and it's great. everyone knows this so i'm not going to repeat it. i however noticed something else more subtle but extremely important - judgment

an orchestrator often has to make judgment calls, like "is this a low-stake change that should be done quickly, or should we really safeguard it", "does this decision need to escalate to the user", "should this task keep going or let's step back". this really tests whether a model was trained with the right rewards

that's where many other models fall apart. sol will remodel my kitchen before making me breakfast. opus 5 will go down rabbit holes and come back with an essay but no rabbit

throughout the whole day, grok 4.6 almost made every decision right for me. i deliberately gave it a lot of autonomy - it merged things when i would have. it asked me for approval for things i actually cared about. it stopped workers from over-engineering. it often recommended ship first fix later when it's safe to do so

chef's kiss

now, the bad:

i have the highest tier supergrok heavy subscription, and it's just not giving enough quota. a full day of mostly single-threaded usage exhausted a whole week of quota. the same work will translate to about 20-30% of the $200 subscription from anthropic or openai using opus or sol

without pricing being on par, i can totally stand behind the model but i can't really recommend the subscription purchase for consumers

overall, grok is on a great trajectory of delivering the most useful model in the world that's fast, pleasant, and capable. i just hope the team realizes how much this pricing problem is holding back their model + grok build harness' adoption, and figure out how to get it fixed

[yantao006 @yantao006]

最高 reasoning effort 已经从 high 调整到 xhigh 了

[yantao006 @yantao006]

正准备睡觉呢，grok 4.6 发布了！升级体验下

[Elon Musk @elonmusk]

Grok 4.6 is now out 🚀🚀🚀

Smart, fast & amazing bang for buck!

[Kun Chen @kunchenguid]

opus 5 is a sign that the obsession with “long-horizon agents” in model training is finally backfiring

i don’t like long-horizon agents, and i’ll explain why they fundamentally don’t work

some people will immediately jump out and say “skill issue”. well, show me one profitable business you built with a long-horizon agent working all by itself - i’d love to learn

so far, the only thing they were able to build that’s even interesting enough for people to talk about are those 3d games that are a partial clone of something that already existed

the reason an agent was able to build a working prototype of complex games like call of duty was that a team of humans already figured out all the requirements years ago for how such games should work, what kind of controls are intuitive, what mechanics are fun etc

all those requirements were already absorbed into the model weights, so when you say “build me call of duty” the model already knows the details. its long horizon execution capability can get all the requirements implemented, which i must say is indeed impressive

but now you can see - the value of long horizon execution has a prerequisite of a massive amount of high quality requirements clearly defined upfront. it took a big team of very talented humans months of effort and many iterations to define that for call of duty

now imagine games like call of duty don’t exist yet, how would we use agents to build it for the first time? we can’t say “build me call of duty” any more. and there’s no way we can define months-worth of game design details upfront

we’ll have to build a tiny prototype of the most basic mechanics, play with it, see if it’s fun, then iterate and expand the complexity. even with the smartest humans, that’s how we work towards something great

we don’t need agents to go dark for a long time, spend tens of thousands of dollars worth of tokens, and come back with a product the agent randomly decided to build - try build something truly novel with this and you’ll see it can’t come up with anything that’s actually profitable (i’ll show you why in a bit)

we need a tight feedback loop where we can collaborate with the agent, plan with it, understand what it’s done, question its approach, apply our judgement, give it real world feedback and iteratively arrive at a good outcome

and that’s exactly what opus 5 absolutely suck at. why? i explained it in more depth with my previous post on how RLVR works - RLVR trains the model to generate code that can pass predefined tests in an isolated environment, which is fundamentally incompatible with the idea of having human in the loop. the more we train the models with RLVR to be “long-horizon”, the less they care about talking to humans

ok now - why do they have to talk to humans? why can’t the models iterate and apply judgement by itself?

maybe one day they could, but not today, due to many limitations. two examples -

1. LLMs today can’t “watch a video” yet. they can look through a lot of screenshots, which is extremely inefficient at observing a high fps animated signal. so anything that requires continuous visual attention is something LLMs can’t do very well

2. LLMs don’t truly understand what’s “intuitive” or “pleasant” for humans. they know what’s already proven to be intuitive and pleasant in the past, but if you present a truly novel concept, it can’t predict whether humans will like it accurately

because of those limitations, human judgment is still needed for almost anything valuable. without humans in the loop, agents will only be able to repeat something that already existed, or go in random directions without true understanding of whether it’s building something useful

in summary, long horizon agents assume requirements all exist upfront. they are fundamentally against human in the loop. and they don’t have true judgement for what humans like

that, my friend, is why they don’t work

[yantao006 @yantao006]

@JamesAI 哈哈，推荐使用下 firstmate (一个开源工具 https://t.co/zpZ1ULOzBt)。
所有项目都通过一个 agent 管理，各个项目的重要信息、个人偏好、全局约束，整体是可以打通管理，后续只需要跟一个 agent 交互，由他来管理各个项目以及分派任务，能够极大减轻多项目管理的心智负担

[在悉尼和稀泥 @JamesAI]

1. 赚有钱人的钱。
2. 找到1这个人群里，你能接触到的群体
3. 在2的人群里，找到和自己学历差，认知差，信息差尽可能大的群体
4. 在3这个人群里，找到最大的共性需求
5. 针对4做标准化方案，不用在乎有没有现有方案，如果有，借鉴或者用自己对4的理解重做一遍。
6. 想办法让5出现在3面前，越多越好

6步走完，发现3不必是互联网圈子，5不用是网站或者产品，6不一定要SEO。

[yantao006 @yantao006]

太感同身受了。不管使用多能干的模型，面对具体的业务问题，再起步阶段也存在一个理解业务、协同的过程。

[Kun Chen @kunchenguid]

sigh.. i have to say something here

imagine you hired a new developer to your company, and on day one he did some terrible work, over-engineering your codebase, speaking jargons all day without context, making all kinds of “genuine mistakes”, lost trust with everyone around him

and then he goes on and tell you - in order for me to do a better job, you need to delete your entire company’s culture and workflow, and have everyone bend over to do things his way, only then can he do reasonable work

oh - and no one tells him how to do his job, NO ONE. he’s always the smartest one in the room and despite doing a terrible job on day one, despite the only results on his resume were vibe coded demos of games that already existed, he demands that you give him a big charter and let him go dark with no communication, taking no feedback

would you have hired a teammate like this?

when humans feel frustrated after using a model, let’s figure out how to RL the model better so they become a better teammate

don’t let a bad model RL _you_

[yantao006 @yantao006]

正如文中所说，广东这边大量的 SMB，意味着其实是有大量潜在的市场供需的，坐标在深圳，利用这个地域优势，走出去，建立更多连接，建立生意人思维，避免工程师思维

[Hawstein @Hawstein]

前两天和一个朋友打电话，聊得非常开心，值得分享一下。

这个朋友是程序员，大小厂都待过。后来实在不想上班，就辞职了。辞职后就想着怎样用 AI 去解决一些具体领域的问题。

为了行文方便，以下用 R 指代这个朋友。

朋友 R 总去一家精酿酒吧，和老板熟了以后，老板知道了 R 在做什么、擅长什么，以及想做什么，于是把他介绍给了自己以前的一位合伙人，下面用 T 指代。

R 和 T 搭上线之后，就一起讨论「AI + 垂直领域」的各种可能性。最后敲定了一个确定性比较强的方向。在这个方向上，R 有技术，而 T 有 SMB 资源。

接下来，R 负责做产品，T 则直接把目标客户拉到旁边，让他们现场看产品做出来的效果，问他们是否满意、是否愿意付费。

和埋头做产品，不知道客户在哪不一样。他们是产品还没做完，真实客户已经站在身边，看着 MVP 输出价值了。

得到 Beta 客户的肯定后，他们接下来准备开始放量了。接下来要担心的，就是量起来后，他们的产品/系统是否能 handle。

他们直接跳过「找不到客户」的烦恼，直接进入下一阶段的烦恼。

和 R 聊天时，我们还得出一个共识：在广东，有大量 SMB，做着一个个细分而垂直的领域。大白话说就是，在这里有大量的小老板，你不会在任何媒体看到他们，但是他们的付费能力足以支撑起你的小生意。

这一点，和欧美市场其实颇为相似。

R 和 T 的故事，是我近期听到的比较喜欢和精彩的故事。

我觉得值得一个 RT。

[yantao006 @yantao006]

给自己的 agent 先注册个 cloudflare 钱包玩玩，哈哈
https://t.co/ZAyiinCE6C.

Reserve yours now at https://t.co/GUz2ytaVtb

[yantao006 @yantao006]

@chunxiangai 小板凳搬好了😀

[yantao006 @yantao006]

@kunchenguid Haha, it really is.

[yantao006 @yantao006]

@Hawstein 现在主要是用 flomo 记一些零碎想法，感觉slidejot 的抽屉式设计的特点是在使用过程中会更便捷？

[yantao006 @yantao006]

@ShawnWen_jp @HiTw93 确实是，看得很细致👍

[Kun Chen @kunchenguid]

if you've been using latest frontier LLMs, it's almost certain that you would have noticed by now the newer models have become worse to talk to

they're more robotic, they speak jargons, they spits out verbose text, and do stuff you didn't ask for

how did that happen? well, i'm not the person who trained those models so i can't speak for certain, but i've known enough evidence that gives me a well-educated guess, and i thought it's interesting to share as a crash course of modern LLM training pipelines

so here we go

let's wind back to 2020. GPT-2 and GPT-3 already came out and were widely available, but they could only predict one token at a time - that's what LLMs are at their core

token prediction was offered via API, but there was nothing you could "talk to". so while it generated a lot of excitement in the academic field due to the emergent intelligence, it didn't have any wide adoption

in 2022, ChatGPT changed all that. the research work that led to ChatGPT was a model initially named "InstructGPT". it took GPT-3 as the intelligent base, and used reinforcement learning with human feedback (RLHF) to teach the models how to "chat"

the core idea of RLHF is that you ask the model to generate a few responses, and then let real humans pick which one they like. do this over and over again, and you get a model that knows how to talk

worth noting even as early as InstructGPT, research found that making the model more pleasant to talk to will reduce their pure academic capabilities. this was called "alignment tax", which is an interesting thing we'll come back to in a bit

there were various techniques done to minimize the reliance on humans, but ultimately the reward is modeled after human preference, making these AI assistants easy to talk to

so remember this: RLHF = training the model to be likable by humans

in 2024, there was an inflection point introduced by claude sonnet 3.5 which was the first model that can kind of autonomously finish coding tasks. it led to the first wave of viable "coding agents"

the way sonnet 3.5 achieved this was by training the model with a harness (now it's called an agent) that has bash and file editing tools, throw the agent into a virtual machine, give it a task, and let it try to complete it. these tasks all have a machine-verifiable outcome predefined, mostly via test cases, that can validate whether the model really finished the task or not

then you let the model do billions and billions of attempts in such virtual environments, and some of them would succeed by chance. you keep the successful agent sessions, and use reinforcement learning to teach the model to do that more, and boom - you get a coding agent

that is called reinforcement learning with verifiable rewards (RLVR). if you look closely, you'll see that in this RLVR process, the final text response from the model doesn't matter AT ALL, as long as the code written by the agent could pass the test. it could talk like a jerk and it would still be rewarded

so remember this: RLVR = training the model to be accepted by machines

late 2024 and early 2025, we saw o1 and deepseek R1 came out as the first wave of "reasoning models". this article is getting long so i'm not diving into reasoning models now, but just know that reasoning models also relied heavily on RLVR to scale the training process - let the model think before taking action, and if the thinking led to a machine verifiable outcome, reward the thinking trace and teach the model to think like that more often

the biggest difference between RLVR and RLHF is that RLVR is more scalable. human feedback is expensive to get, especially in domains where only an expert can have a valid opinion on which result is good

with RLHF, if we let the model generate 100 responses, then a human has to review all 100 responses to pick which is good

with RLVR, the human (or sometimes an AI) would define a task and verifier only once, and the model can generate a million responses - the machine verifier will pick which responses are good in an automated way

so as a result, RLVR is becoming more and more dominant in newer models' training pipeline

if you put all these things together:
- RLHF = training the model to be likable by humans
- RLVR = training the model to be accepted by machines
- RLVR is more scalable
- "alignment tax" says "likable by humans" makes the model do worse on verifiable tasks

now you see why the newer models are becoming less and less likable?

this is not just a "frontier labs screwed up their model training" problem - this is a war between machines and humanity, and humanity is losing

we chased after benchmarks, when none of the benchmarks measure whether humans actually enjoy working with the model

we use machines to decide which AI response is better because that's easier and cheaper, when we have no way of making sure those machines actually represent what we humans want

we let AI go dark in a virtual environment on its own and complete predefined tasks at all costs, when in reality we often cannot define a verifiable outcome upfront, and need AI to work with us along the way

i don't have a good solution to this, but i want to call for awareness that we're starting to witness a failure in aligning super intelligence right in front of our eyes

this war between machines vs humanity is one we really can't afford to lose

[yantao006 @yantao006]

最近一个新站本来准备接 creem 的，刚好看到小排@bourneliu66 的公众号文章，果断尝试了 waffo，接入体验很顺滑，很快就审批通过了。

附小排大佬的邀请链接：https://t.co/p7jrADNjID
- 1000 USD交易免手续费(26年底有效)
- 审核绿色通道(1个工作日内完成初审或反馈)
- 社群客服(有问题可以随时在微信社群内提问 会有专人解答)

https://t.co/bj1Gn3G7Se

[yantao006 @yantao006]

@blankwebdev 感谢大佬，太及时了，作为新手刚开始 SEO学习和实践，做关键词研究、网站分析、竞品分析，这套 api 能力正好可以用上👍👍

[yantao006 @yantao006]

太及时了，作为新手刚开始 SEO学习和实践，做关键词研究、网站分析、竞品分析，这套 api 能力正好可以用上👍👍

[blank @blankwebdev]

🎉 新项目 https://t.co/UjVra4x2Tb 上线，送一波积分作为福利。

🎁 上线福利：每人赠送 5,000 积分，本次赠送的积分有效期为 1 年。

参与方式：

✅ 评论这条推文
✅ 引用或转发这条推文

TabAPI 把网站流量、WHOIS、RDAP、DNS、外链、反查 AdSense、Google SERP、网页转 Markdown 和网站截图等能力，整合进了一个 API。

不用注册多个平台，也不用维护一堆 Token，一个 Token 就能调用，直接返回干净的 JSON。购买积分无需订阅，一次购买，永久有效。

领取方式：完成以上步骤后，私信我你的 TabAPI 账号，会按照顺序发放积分。每个账号限领一次。

活动截止：8 月 7 日 23:59:59

[yantao006 @yantao006]

@Alex_Yuetmou 用 firstmate 作整体项目管理，可以在一个项目下根据不同的需求调度不同大模型，同时做单项目和跨项目的经验打通。这个调度和编排理念是比较超前的，有时间可以看下项目介绍～

[yantao006 @yantao006]

最近一直是把 Grok 4.5 作为主力模型，也已经彻底把 Firstmate 切换到了 Grok 4.5。

甚至对于 crewmate ，在把模型从 gpt5.6 切到 grok4.5 之后，都有点不想切换回去了。

不同模型带来的使用体验非常微妙，有时很难准确描述这些细微的差异，但你就是能够感受到：自己更愿意、也更倾向于使用哪个模型。

再次推荐，我认为理念上可称为next-level 的 ai agent 协作工具：https://t.co/odqvBokixd

[Kun Chen @kunchenguid]

now that the dust is settling around the last wave of model releases, and i've had enough time to use all these models in practice, let me share a more complete set of thoughts

1. grok 4.5 is probably the single most significant event during the last couple of weeks

i've been talking with many heavy users across model families, and it's pretty much a consensus that grok 4.5 is the most "pleasant" frontier model to work with day to day. it's incredible how precisely the team behind it found this perfect sweet spot and created a model that's so fast, efficient and capable

it proved spacexai is now a 3rd real player in addition to anthropic and openai. they have real time data from the biggest public townsquare of humans, they have acquired a popular agent harness, and now they have proven they can build great models. and if you look closely, they are designing their own chips, they have their own data centers, they can send GPUs into the space and create tokens out of sunshine

holy shit

2. speaking of data, human usage over a harness proved to be extremely important for training good models. grok 4.5 was the first model that incorporated cursor's data and it made a massive difference compared to previous generations of grok

this explained why amazon mandates employee usage of kiro, why meta installed mass surveillance over employee devices, why anthropic bans 3rd party harnesses, and why google is still struggling with gemini - because they don't have a popular harness with mass adoption to collect the data

this is part of why i don't think the subsidized LLM subscriptions will end any time soon, because a wide consumer adoption is the best source of data collection. we're paying the subsidized tokens by teaching their models how to get work done

3. opus 5 flopped. almost no one likes it. the only people who like it seem to be using it to one-shot 3d games that look impressive but no one will ever buy

anything that AI can one-shot is just the definition of garbage, because if you can one-shot this thing with a quick prompt, you should know that it means billions of other people can also do it - you will not create anything of value this way

and this is just a symptom of a more fundamental problem that model training is heading down a slippery slope where machine verifiable outcome is dominating over human feedback

the latest training process rewards the agent for running for a long time and finishing a complex project, yet no longer seems to care about how the agent talks to its human

jargons, walls of text, "an honest mistake" - opus 5 showed us that we need AI that's more human friendly. let's not build a world where we end up working with robotic a**holes all day

4. fable 5 remains undefeated as the upperbound

i talked about this in my previous post about wisdom vs diligence. the benchmarks blend both together so it's not easy to see, but fable 5 is the GOAT on the "wisdom" dimension despite it not winning on every benchmark. if you used it meaningfully, you know what i'm talking about

that said, it seems anthropic is extremely paranoid about other players, including open models, reaching the same level of intelligence, which is an indication that the moat is not strong. kimi k3 is just a preview of what it looks like

at the same time, fable is the first time token cost is becoming a very real problem. it's the only model so far that i can't afford to keep using all day. i suspect this will remain true for a while, that we have to pick and choose what tasks to give to fable-tier models, not using them as a daily driver

5. openai is in an interesting position

gpt models have been great at efficiency, but now grok is also very competitive. gpt also haven't quite reached the same wisdom upperbound where fable is yet - although gpt 6 may change that

i think there are two angles for openai to pursue:

- continue to bet on efficiency, and go after enterprise adoption while claude is too expensive and grok has a brand tax to pay there. this is a very viable strategy

- or.. compete head on with fable on wisdom and win against them on the "human friendly" aspect. although traditionally this hasn't been the strength of openai models either, so this feels like a low-ROI option

alright, that's a bit of a long post but the landscape is just becoming increasingly complex. hope these thoughts are helpful in terms of providing a reference for how to rationalize everything happening

[yantao006 @yantao006]

@ccbuhejiu 是的，我订阅了 grok heavy，现在有折扣优惠，感觉还不错

[yantao006 @yantao006]

@stevenmoon2460 哈哈，强烈推荐~ 开源贡献者是个技术大牛