Jason Erickson (@jasonthink) header
Jason Erickson (@jasonthink) profile picture

Platform Engineering lead. Practical AI and tools that make work lighter. Writing on managing time, AI at Work & platform eng. Mentor on MentorCruise. Dad of 2.

Southside, TN
Joined October 2016
Photo by Jason Erickson (@jasonthink)Photo by Jason Erickson (@jasonthink)Photo by Jason Erickson (@jasonthink)Photo by Jason Erickson (@jasonthink)Photo by Jason Erickson (@jasonthink)Photo by Jason Erickson (@jasonthink)Photo by Jason Erickson (@jasonthink)Photo by Jason Erickson (@jasonthink)Photo by Jason Erickson (@jasonthink)Photo by Jason Erickson (@jasonthink)
Jason Erickson retweeted
Karpathy 开炮了。 这位前 OpenAI 联合创始人、前特斯拉 AI 负责人,在一段 16 分钟的对谈里,把当下 Agent 热潮的底裤扒了个干净。 他说得很直接:Agent 的本质就是大规模蒸馏。你花大价钱跑的那个大模型,99.99% 的容量都浪费在了它压根用不上的数据上。听起来刺耳,但你仔细想想,事实就是这样。 然后他给出了一条真正能跑通的技术路线:先用大模型把能力蒸馏到小模型里,再给小模型配上正确的工具,最后用一个闭环把整个系统串起来。小模型、正确的工具、一个闭环,三件事,够了。 这里面最关键的一环,是闭环。 小模型本身并不可怕,可怕的是它一旦接入闭环,就能自我校正、自我改进,越跑越强。模型的参数量可以小,但只要反馈回路搭对了,它就能长出远超参数规模的实战能力。 所以别再为更大的模型交税了。真正让一个 Agent 变得危险的,从来都不是参数量,是闭环。 这 16 分钟的信息密度,比今年市面上任何一门 500 美元的 AI 课都高。看完之后,再去读那篇 Harness, Loops, Graphs,你就会明白:闭环就是那副马具,套上了,小模型才能真正跑起来。
62
94
12
732
119,776
Yep.
you cannot automate what you do not understand and to understand, you cannot first automate slow down to go faster i guess
9
My bingo card is full this week. I'm going home.
8
Be different...
You have to do different things to get different outcomes....and no one wants to be different.
1
3
80
It's important to savor the little victories. I try to think of 3 good things every day, no matter how small. It helps.
Savor the little victories as much as you criticize the little mistakes.
11
I once asked AI for account information and got a next step of “I will reboot the machine after I remove the root volume.” That misfire taught me to focus on the skill's When. Task, context, trigger. jasonthink.substack.com/p/wh… #AI
13
When, How, What: Three Keys to Agent Skills jasonthink.substack.com/p/wh…
14
One of my lead developers had a worthy #QuoteOfTheDay "I don't think I need AI to order a burrito" LOL
1
1
24
Thanks. I've learned in today's corporate world, staying visible and always moving forward is critical.
3
I have to look into this.
I haven't seen another harness work like this, infinitely extensible, a real working system, and built for power users.
2
46
Hmm. Not a bad idea. I wonder if I can.
My mental health improved significantly when I started leaving my phone at home. Often. As a default. Go to the gym without it. Go to dinner without it. Take a walk without it. Go to a museum without it. At first, you will start to come up with justifications for why you can’t do this. But once you overcome them, you will start to experience a wonderful mix of creativity & happiness as you unshackle yourself from your handheld dopamine casino.
1
4
990
Why do we still have issues with Sweet N Low? Hasn’t this science been settled.
28
Do you remember when you joined X? I do! (I don’t actually) #MyXAnniversary
1
52
So many slack channels. So many business processes have moved to slack channels. Now need a directory to find slack channels. <- This has become the new email jail and I don't like it.
14
I stress a lot when I see a problem in the future. I warn my team but they don't see it and continue on. I could be wrong but I don't feel wrong. I trust my gut. I don't know what to do...so I just go with the flow until it comes true.
20
Jason Erickson retweeted
"Let’s plan the roadmap for next quarter” is the modern day equivalent of “let’s fax everyone the strategy.” We're going into q4 planning and it makes my brain hurt. Roadmaps existed because software was expensive to build. You needed a bunch of people. Meetings. Someone moving little boxes around Jira for three months from now. It makes no sense and I'm trying to work out what we call the replacement. Maybe just "Continuous Compilation", ideas just compile to working product. I want a list of ideas that we want to continuously compile. That's the roadmap now. The cheapest way to figure out whether something is a good idea is often to just build the fucking thing. Here’s how I work now. I've done this weekly over the last few months and it works. My current napkin math is that I can usually build 6 months of old-school roadmap in a week. Sometimes more. I've also hit a wall sometimes. It happens. When we have an idea, I usually 10x the ambition and then write a prompt, not more than a page long, that basically says: - here’s the ambitious goal - here are a few use cases - here’s roughly how I think we should build it in terms of where it fits into code base, design considerations, etc.. - here’s how we’ll know if it actually works - now go fucking surprise me Agents build all night. Not even with a crazy factory or complex agent or agents. The base models are just too good for that now. In the morning I look at it. Ok, they usually build in a few hours. Usually some of it is great. Some of it is stupid. Some of it exposes that my original idea was wrong. And almost every time it gives me 5 new ideas I couldn’t have planned upfront. So we throw shit out. Who cares? We learned something. Then we change direction and send them off again. It usually takes 3 rounds before the shape appears and we've learned about the gremlins and possibilities. 95% of the time I end up building something with 10x the scope of what somebody had carefully spread across the next quarter in Jira. “What should we put on the roadmap for next quarter?” My brain now translates that to: “Oh, you're stuck in the 1500s”
33
19
8
345
40,758
Every week, the average professional attends 23 hours of meetings. That’s nearly three full workdays — gone — with no output to show for it. I can help you fix it. Read more: jasonthink.substack.com/p/wh…
17