Thorsten Ball
@thorstenball
Followers
35K
Following
19K
Media
3K
Statuses
22K
Author of https://t.co/arW0hnVET0 and https://t.co/RN9xXOzhON. @sourcegraph working on @ampcode. Ex-@zeddotdev. Programming where the rubber hits the road.
Großwallstdt, Bavaria, Germany
Joined November 2011
My second book "Writing A Compiler In Go" is now available!. You can buy it here as a single eBook or in a bundle with its predecessor."Writing An Interpreter In Go":. #golang.
compilerbook.com
This is the sequel to Writing An Interpreter In Go and this time we're writing a compiler and a virtual machine for Monkey. Same codebase, same approach, new goals. Code front and center, step by...
40
202
920
What didn't fit in: (3) is still very fast. I use it a lot. A lot.
0
0
2
So far, my 3 phases of using agents:. 1. slot machine high. "i have endless powers now". try it on everything. 2. grand plans. huge features, multi-step elaborate flows, checklists, plans. 3. reality. targeted edits i'd know how to make. prototypes. hail marys. explore & learn.
3
3
19
Model alloys, man. (I guess: release #4 today).
Raising An Agent - Episode 8. In this episode, @beyang sits down with @camden_cheek to discuss how the Amp team evaluates new models: why tool calling is the key differentiator, how open models like K2 and Qwen stack up, what GPT-5 changes, and how qualitative “vibe checks” often
0
1
13
Finally good news for people who have to write Java. (I'm kidding! Love me some JVM)
Amp CLI can now read diagnostics from JetBrains IDEs for JVM languages such as Java, Kotlin, and Scala.
0
0
14
damn it, that's perfect and I wish I had put this in the demo.
If you work with GitHub daily, I highly recommend using this with. gh issue view 1110. And again with the comments flag. Likewise for PRs. And you can even grab code review comments for PRs too!. Massive productivity boost.
0
0
11
You can now run shell commands directly in the Amp CLI. (Finally!)
8
9
89
So apparently we have a :corndog: in Slack and it has both ketchup *and* mustard on it?!
9
0
21
Surprise #3: I feel dumb *and* very cool at the same time when I pull over the microphone before talking to the agent. "Alright, listen, here's what I want you to do. ".
Trying speech-to-text heavily this week and biggest surprise so far: surprisingly little amount of text produced after talking for 3-4 minutes.
3
0
25
Amp Tab was so good today. Here's what I shared in our internal channel this morning. Little moments of delight, it's neat.
Amp Tab now uses a new, custom model to provide better completions. We rebuilt our training dataset with synthetic examples and used DPO post-training on top of SFT. The model learns from its previous bad outputs as negative examples alongside synthetic ground truth as positive
3
1
22
Surprise #2: when I now talk to the agent using speech-to-text I somehow end up saying "Thank you." at the end of my prompt, as if I was leaving a voice mail. Very dumb.
Trying speech-to-text heavily this week and biggest surprise so far: surprisingly little amount of text produced after talking for 3-4 minutes.
6
0
22
"I would just simply make the whole thing out of the oracle"
1
0
11
Ugh, I meant Bose QuietComfort Ultra Earbuds. The naming, man.
0
0
0
Mini tech-review:. Okay, AirPods are good. I had the Bose QuietComfort Ultra before and really enjoyed them, but the *one* thing that made me mad was that you would *lightly* brush them, say if you touch your hair/hat/sunglasses and they'd register that. AirPods fix that for me.
4
0
11
Let's do it live :).
Get ready for Next Token!. New live show every Wednesday hosted by me and @thorstenball. Launching Sep 3rd. Guests already confirmed: .@dhh .@mckaywrigley . If you're an engineer using agents, you won't want to miss it. @theo and @ThePrimeagen - want to hop on the show
4
1
26
Trying speech-to-text heavily this week and biggest surprise so far: surprisingly little amount of text produced after talking for 3-4 minutes.
3
0
35
Okay, this was pretty magical. I was debugging an issue, hand-in-hand with the main agent: adding logs, trying stuff, asking the agent, sharing more info. Then, once I kinda knew where the issue was, I asked main agent to ask the oracle for what the issue is. It found it.
10
6
96