Beyang

7.1K posts

Beyang banner
Beyang

Beyang

@beyang

Amplifying human agency @ampcode, founder @sourcegraph

San Francisco, CA Katılım Eylül 2008
2K Takip Edilen12.6K Takipçiler
Beyang
Beyang@beyang·
AI ads these days be like
English
1
1
10
698
Beyang
Beyang@beyang·
@a3b0_ Lol this is my favorite one
English
1
0
1
26
a3b0
a3b0@a3b0_·
@beyang call it Roy (ie:orbison)
English
2
0
2
36
Beyang
Beyang@beyang·
We're trying to figure out a name for a new agent in Amp. Shot down so far: Orbert, Orberford, Orby, Borb.

(We like Orbs, can you tell?)
 No one likes my idea, but I maintain that if AI is like nukes, like the Safety People argue, the only proper name is Orbenheimer.
English
44
0
46
4.3K
Beyang
Beyang@beyang·
@kenryu42 Is that “orb” in another language?
English
1
0
0
102
latent space marine
latent space marine@xlatentspace·
@beyang I have not observed an output quality difference but I do miss my old collaborator I said nothing because I trust you guys
English
1
0
1
69
Beyang
Beyang@beyang·
It's funny, the combo of recent model releases + some changes in Amp defaults means that our token load has quite recently shifted from 80/20 Model Provider A to 20/80 Model Provider B. Nary a peep from users about this change. Has something fundamental shifted? Maybe.
Beyang tweet media
English
4
0
16
1.2K
奇怪的猫
奇怪的猫@astrangecat0·
@beyang Wait, can you add/remove labels directly from the command palette without opening anything else?
English
1
0
0
41
Beyang
Beyang@beyang·
Amp Orbs make it so easy to fire off threads that now users are asking for better thread management tools. So we shipped thread labels with the ability to add/remove/delete via command palette. Way easier than juggling 5 different agent/app windows/workspaces/tmux sessions.
English
7
2
52
3.1K
Beyang
Beyang@beyang·
This has effectively become my personal issue tracker / todo list.
Beyang tweet media
English
5
0
55
8K
Beyang
Beyang@beyang·
@dilika @threepointone To install raw models, you need to install raw model agents via plugin into your repository, like this: amp plugins add --target $YOUR_WORKSPACE --auto-update @amp/grok-45-mode
English
1
0
1
297
sunil pai
sunil pai@threepointone·
steve jobs would’ve hated model selector drop downs and the discourse around comparing versions and thinking levels
English
84
55
2K
222.2K
Beyang
Beyang@beyang·
@wan_ruirui We only use Gemini 3 and Haiku for minor tasks now. The main agents and major subagents are driven by newer models: ampcode.com/models
English
1
0
0
64
Ruirui Wan
Ruirui Wan@wan_ruirui·
@beyang Any updates on using gemini3 haiku4.5 things like that to the newer version of the model? Feels these model getting old now
English
1
0
0
58
Beyang
Beyang@beyang·
Grok 4.5 in Amp. It's not yet tuned and optimized, but we invite you to experiment with us. If you haven't yet installed Amp: curl -fsSL ampcode.com/install.sh | bash Then try it out: amp update amp plugins add @amp/grok-45-mode --auto-update amp --mode grok45
Beyang tweet media
English
23
18
132
584K
Nacho Britos
Nacho Britos@NachoBritos12·
@beyang How do I know which LLM is being use in low, med, high?
English
1
0
0
77
Beyang
Beyang@beyang·
Agents != Models I've said this before, but it bears repeating: Amp's agent modes take away a lot of cognitive load. A user sent me a screenshot, I pasted it into low mode. "Wait, can that model do images?" The model can't, but the agent was fine:
Beyang tweet media
English
4
2
42
3.3K
Thorsten Ball
Thorsten Ball@thorstenball·
Someone on here said yesterday "Stop trying to make orb happen, it's not happening" But I just walked past someone on the street who said "Time to get orbin'!" And then they spawned seven Amp agents in orbs. So, it is happening. (I just made that up, it's not real)
English
22
0
92
12.4K
Beyang
Beyang@beyang·
Protips on harness engineering
Hugo Bowne-Anderson@hugobowne

Yo @nicolaygerold just rebuilt @AmpCode's enterprise coding harness for Fable 5 and GPT-5.6. He needs AI builders to know this: kill your own harness before the next model does it for you. Every component in a harness encodes an assumption about what the model cannot do itself. Those assumptions expire. AMP had an experimental plan mode. When one instruction could make the model plan without editing, Nico ripped the feature out. Yesterday, Nico and I taught a two-hour workshop on how coding agents are actually built, and how their harnesses must be rebuilt as models improve. Nico takes us inside the core loop, tool schemas, context engineering, AGENTS.md, compaction, handoffs, sub-agents, Bash, Apply Patch, and the decisions that go into adapting a production harness to a new generation of models. If you build agents, this is a rare look inside the work of someone rebuilding an enterprise coding harness at the frontier. Go get it ↓

English
0
0
19
3.3K

Keşfet

@a3b0_ @kenryu42 @dylayed @xlatentspace @astrangecat0 @dilika @threepointone @wan_ruirui