X / Twitter
If your eval system says you are achieving 100% w/AI accuracy then your product is actually deeply broken, or you are tracking an irrelevant set of metrics
Remember, don't raw dog your prompts
build evals (or you'll lose your mind like me) https://t.co/cvym3JhL4a
Web UI for Apple Depth-Pro
Metric depth estimation determines real-world distances to objects in a scene from images. This repo provides a web UI that allows users to estimate metric depth and visualize depth maps by easily uploading images using the Depth Pro model through a simple Gradio UI interface.
This is pretty cool!
It creates a RESTful API around the new SQLite storage inside Durable Objects and gives you a fully working SQLite database that you can interact with over HTTP.
They also cooked with their landing page https://t.co/PAHXSMg4WF https://t.co/NqK6gyBC3U
I’ve always bought the “writing before reading” approach to literacy on the merits of general arguments and evidence, but seeing it play out with my child, it is even better than I thought it would be.
My 3 management non-negotiables https://t.co/33JzMPbzPm
The Agent Economy explained:
While makers are busy cloning Product Hunt or other successful projects (I counted 7 clones for my projects in September alone), the real "BIG NEW THING" is "Agents."
I get paid $2k in 6 months for an agent compared to $9 SaaS.
Why? Agents replace actual employees.
Something we, the business people, have been dreaming in our wet dreams.
E.g. I had a marketing department that I paid $10k/mo for and now it's 99% covered by seobot agent with a little human moderation.
What am I willing to pay for such agent? At least 10% of what I paid humans. So $1k/mo is a no-brainer.
Would I pay $1k/mo for SaaS? Ofc no. I pay less than that for 40 saas tools combined(most used for free).
- Agents scale horizontally.
I run agents for my directories. Listingbott to get backlinks and seobot for blog. But I run not one, but 30 directories. If I had humans employed, I'd have to hire new employees every time I start a new directory. Hiring is really hard and expensive.
With agents, I simply scale them horizontally and pay more for the usage. Just one click. No hustle.
- Agents get smarter.
It's difficult to teach humans. It takes time and those a typical startup can afford to hire are usually not so capable of learning.
Agents learn new things with every new release.
Every week or months they are capable of more.
So every instance of the agent in every product I have suddenly gets smarter after release.
- I can pause agents.
It's difficult to hire and fire employees. So sometimes you dont need them, but you keep them because you dont wanna fire them and later spend time looking for new hires again.
With agents, I can turn them off if not needed and then turn back on any time.
The Agent Economy:
- only 0.0001% of agents have been created. We need 50,000 more agents in the next 5 years.
- it's more difficult for corporations to make them, because of the risks. Indie makers can take more risks, which means we have an advantage.
Two of my agents (please don't clone them 🙏):
@seobotai - blog articles, internal and external linking, keyword research, scrpaing, proofreading, no hallucinations, images, videos....
@listingbott - submits to all existing directories (100 more relevant) (10k in the db and growing), brings backlinks on autopilot and boosts domain rating to win more SEO traffic.
https://t.co/QmhrH0xrBz
Having a blast working on a tool to create wide gamut color palettes for @tailwindcss v4. https://t.co/HER3yIIkvd
How do you do move activation?
Break it down in to setup, aha, and habit: https://t.co/AwhLHePZs3
notes from el presidente @pirroh on building agents during langchain x @latentspacepod devday pregame:
- “dozens” of calls per agent
- use ops tools. dont rawdog llm apis lmao. they use langsmith ofc
- claude sonnet ftw
- dont use single agent. have manager agent supervise 2-3 agents. any more and its hard to debug
- ship small product wins rather than big project
- accuracy > cost > latency
- betting on cost going down over time (aka replit agents is bleeding money rn)
- evals are hard
- surprisingly mobile app use quite high (maybe not surprising since need to type less)
- dont use self consistency. not worth it
- @altryne had good qtn on balance between teaching beginners and being power tool for power users
I and @DennisSmolek just updated the React wrapper for @garrettkjohnson's #threejs Path Tracer.
Path Tracing in #r3f is as simple as:
<Canvas>
<Pathtracer>
<Scene />
Check it out: https://t.co/SC9oZMklIu
Demo: https://t.co/6WQTGh6ZiE https://t.co/MkYtWK1Xjf