Beyang @beyang

Amplifying human agency @ampcode, founder @sourcegraph San Francisco, CA Joined September 2008

Tweets

7K
Followers

12K
Following

2K
Likes

19K

Beyang @beyang

16 hours ago

@sixhobbits Fair enough, I suppose those are good priors. Makes it confusing to talk about human agency and AI agents, though. Never a better time to highlight the contrast!

0 0 0 6 0

View Details

If you think about it, "agents" is actually a pretty bad term, perhaps coined by folks who view AI as a human replacer rather than an amplifier. AI agents have no actual agency. The agency comes from humans and is irreplaceable.

15 6 66 2K 3

View Details

Beyang @beyang

16 hours ago

@traskjd Fair enough, I forgot about that use of the word!

0 0 2 90 0

View Details

Beyang @beyang

23 hours ago

@abacaj x.com/beyang/status/…

Beyang @beyang

3 days ago

Have been running this all weekend. It’s been awesome. FWIW still not competitive with GPT 5.5 in Amp’s deep mode on longer horizon or more nuanced tasks, but it’s fast and robust and great for targeted stuff.

8 1 77 9K 15

0 0 7 1K 0

View Details

Beyang @beyang

23 hours ago

@xeophon Yeah seeing if we can prompt this behavior out.

0 0 1 128 0

View Details

Beyang @beyang

2 days ago

Here's an example where I asked both GLM 5.2 in Amp and Amp's deep^2 mode (GPT 5.5 medium thinking) to fix a flaky test. I then asked deep^2 to grade GLM's solution against its own. Both got to comparable solutions, but deep^2 preferred aspects of GLM's solution. The cost delta: - $1.49 for GLM - $5.02 for deep^2 But here's another facet: deep^2 actually completed the task significantly faster than GLM. Despite being faster on a tokens-per-second basis, GLM often emits many more thinking tokens and so takes longer to arrive at the end of a turn. So to recap: for this simple task, the quality was comparable, GLM was cheaper, but deep^2 was faster. So which should you use? There's no question that there is alpha you can extract from using GLM 5.2 in specific cases. But to put the onus of that choice on the user is a mistake IMO. The cognitive load of these meta-factors greatly outweighs the mental energy you'd want to expend on this task. It's the meme of the AI engineers that feel extremely busy and overburdened, but aren't actually shipping at a higher rate. As the agent-capable model landscape explodes, I think you'll see more recognition of this cognitive overload—and more people reaching for solutions that relieve them from the burden of choice for everyday tasks.

10 3 95 7K 24

View Details

Beyang @beyang

23 hours ago

@harsh_thakur_1 I didn’t track precisely but it was noticeable. At least a couple minutes I’d say.

0 0 0 40 0

View Details

Beyang @beyang

2 days ago

@0xV0LYX I read the code in this case and agreed there was something better

0 0 2 226 0

View Details

Beyang @beyang

2 days ago

@xlatentspace 100% agree, but... maybe I'm working on 10 things at once because I need to wait. The correlation/causation is difficult to disentangle here.

1 0 2 68 0

View Details

Beyang @beyang

2 days ago

A lot of people going for smart main and fast subs. The intuition there is the main agent is sort of the "brains" of the operation with dumber "minions". But another way to think about it is that the agent closest to the human should be the speediest, since waiting sucks!

Beyang @beyang

4 days ago

Which do you think you’d prefer? Fast main agent with smart subagents? Or smart main agent with fast subagents?

28 0 40 9K 4

10 1 30 3K 2

View Details

Beyang @beyang

2 days ago

A simple Amp plugin that codifies an interactive-plan-async-implement way of working with agents:

Tim Culverhouse @rockorager

5 days ago

Along with custom agents, we've also made it possible to spawn new threads with the builtin Amp agents. It has massively changed how I work with agents. Here's how: My core workflow is to have a large thread where I'm iterating on a design. Traditionally, I've dumped this to

3 3 58 14K 59

1 1 39 6K 18

View Details

Beyang @beyang

2 days ago

@benvargas Should work now

1 0 1 48 0

View Details

Beyang @beyang

5 days ago

GLM 5.2 beats Fable on this visual design benchmark, but not being a multimodal model, it can't review its own work. So we gave it a pair of agentic eyes in Amp.

Design Arena @Designarena

5 days ago

x.com/i/article/2067…

72 234 2K 1.6M 2K

7 10 154 22K 76

View Details

Beyang @beyang

2 days ago

Safety is the banner for invasion of privacy and restriction of freedom.

Quinn Slack @sqs

3 days ago

Every human has Freedom of Intelligence. Nobody & no gov't should infringe on our right to use available intelligence. Companies shouldn't be compelled to offer models freely, but they should be shunned if they engage in intelligence censorship, esp. if they lobby gov't for it.

11 8 110 16K 6

4 2 25 2K 2

View Details

Beyang @beyang

2 days ago

@RicheTechGuy Can you share some threads via DM? We fixed a few issues over the weekend.

1 0 1 53 0

View Details

Beyang @beyang

3 days ago

Beyang @beyang

5 days ago

GLM 5.2 in Amp. Zero data retention, US and West-based inference. It's a good model. Try it now: amp update amp plugins add --auto-update @amp/glm-52-mode amp --mode glm-5.2

58 48 673 103K 426

8 1 77 9K 15

View Details

Beyang @beyang

2 days ago

@tcballard @rockorager probably the best person to answer this

1 0 2 41 0

View Details

Beyang @beyang

2 days ago

@traskjd For threads where both work, the cost difference is 5-10x. Where glm doesn’t work it obviously doesn’t work but the cost is low enough in most cases I’m only out less than a dollar. You do pay in wasted time, though, of course

0 0 2 77 0

View Details

Beyang @beyang

2 days ago

@0xV0LYX Yep

0 0 0 89 0

View Details

Quinn Slack @sqs

3 days ago

Should government serve the interests of the steel industry, or the interests of everyone that uses steel? It's the classic problem of concentrated benefits and diffused costs (fee.org/articles/conce…).

Beyang @beyang

5 days ago

With GLM 5.2 closing the gap between open and closed models, the knee-jerk reaction in Washington on both sides of the aisle will be to try to limit access to Chinese models. But we need to recognize that the American AI ecosystem has many actors with competing incentives and

7 7 129 15K 19

1 2 44 8K 4

View Details

Beyang @beyang

3 days ago

TLDR yes But with one nuance: registered agent modes are mainly for letting users start threads in that mode. But a plugin can expose the same custom agent through a tool, and then the primary agent can call that tool to run it as a sub-agent/child thread (using parentThreadID) or even kick off a separate top-level thread. Examples here: - ampcode.com/manual/plugin-… - ampcode.com/manual/plugin-…

0 0 0 36 0

View Details

Beyang @beyang

5 days ago

It's now super easy to create specialized agents through an Amp plugin. Select the model, define the system prompt, and just inline the tool definitions. This is all the code you need (and you can have Amp write it):

Tim Culverhouse @rockorager

5 days ago

Custom Agents are now available in Amp via plugins. Create custom agents or subagents, create threads and tools with them, and perhaps most fun: send messages agent-to-agent.