Quinn Slack

CEO & Member of Technical Staff @ampcode, founded @sourcegraph

[Quinn Slack @sqs]

@jhicks154 If you have a ChatGPT sub active in Amp, then it’ll be used for rush and deep modes. It passes along your speed (fast or regular) but I am not sure how their endpoint handles it for subs.

[Quinn Slack @sqs]

spotted this little guy in Amp

[Quinn Slack @sqs]

@MichaelRyerse Thanks!

[Quinn Slack @sqs]

The Factory team seems very smart, and I'm eager to see how well their cost-optimizing new model router works. If they can pull it off, that is a huge accomplishment.

For Amp: we may try to build a model router that picks the best model for a task, but we don't intend to build a cost-optimizing model router based on the current state of the models. Here's why.

Every time we've looked into using cheaper models in Amp, we've benchmarked on tasks that reflect how people use agents for coding today. On these real tasks, the expensive frontier model was not only the best (obviously), but also usually the fastest and cheapest, when measuring end-to-end task completion.

Why? Cheaper-per-token models are less capable, which means that on complex real-world tasks they spend more tokens and time fixing mistakes along the way.

You can find plenty of cases where cheaper models are indeed faster and cheaper end-to-end. But such cases were rarer than we expected, and the differences were fairly small.

If you can easily detect such cases, then there is an opportunity here. But even then, on the AI hedonic treadmill, once people get a taste of frontier intelligence, they don't want to go back to using those more primitive prompts where cheaper models suffice. (Which is a good part of human behavior! It's how we decided to stop living in caves!)

If your tasks can be handled just as well by non-frontier models, I would strongly advise you to uplevel how you use agents and what you produce to stay competitive against people who are using frontier models.

In a power-law world, with rapid intelligence advances, try to get to the frontier and stay there.

[Quinn Slack @sqs]

@MichaelRyerse Where?

[Quinn Slack @sqs]

@NilkhanVaibhav @FactoryAI Awesome, let me know how you like it

[Quinn Slack @sqs]

No plans, but that could make sense. Not sure what to do with Amp Free now. Our goal is to make the frontier more accessible to more people so we can explore together, not to subsidize routine usage, although I know that is very helpful to many people. The subsidy is smaller and smaller compared to spend anyway.

[Quinn Slack @sqs]

@JonathanHaas Yeah. The counterargument is that it's possible to build a classifier to detect which prompts /can/ be solved easily by a less capable model. I can see that having some value, but not nearly as much as people think.

[Quinn Slack @sqs]

@gojun077 @united Awesome!

[Quinn Slack @sqs]

The final boss of an interactive client<->server application is intermittent airplane wifi.

Amp worked pretty smoothly for me on a 15-hour SFO-SYD @united flight, but we want to make it buttery smooth.

If you have a flight in the next 10 days and want to help us test, reply here with your Amp username and flight route, and I'll send $250 in free credits. You then just need to post your feedback using Amp in flight. Would also love a photo of Amp in flight, and a comparison to how other agents work in flight.

(I'll have Amp decide for me which of the replies seem legit. I reserve the right to not pick everyone.)

[Quinn Slack @sqs]

Despite being on airplane wifi, the Amp CLI's new architecture (running the agent loop on the server) let me ship some top requests today because it sends ~95% less data to/from the server.

• Alt+t/Opt+t shows thinking
• Cost is shown, but it's cooler if you click to toggle to cumulative tokens

(If you missed it: https://t.co/QwbrOIMuDK.)

[Quinn Slack @sqs]

@bytebot It’s one of the reasons I’m nervous about subs. We want Amp to always always work the best, and when that’s not true, we want it to be within our power to fix.

[Quinn Slack @sqs]

@bytebot Yeah, I’d expect subs to be slower since Amp’s own model access is heavily monitored and optimized. We can’t do the same for subs.

[Tim Culverhouse @rockorager]

Those of you who are not fortunate enough to work with Thorsten may not know this: when you work with him, you get an extra, private newsletter in addition to the public one.

Every. Week.

Think about that.

[Quinn Slack @sqs]

@aiandcloud @rockorager @Copilot I'm seeing it point to ampagent, which is correct. That is the new name of the Amp GitHub user. What's it showing for you now?

[Quinn Slack @sqs]

@HankYeomans It feels so freeing. I never need to worry about someone using Amp in a way that I’d feel bad about guiding them to. I highly recommend it to others.

[Quinn Slack @sqs]

I love when technological progress kills a feature.

The models are so good that (A) this feature is too risky and (B) it's not needed anymore to convince people that agents are good.

https://t.co/PlPosB3wYu

[Quinn Slack @sqs]

@lebotchine Yeah, but another way to look at it is that Microsoft was subsidizing you heavily. You should complain about all the other companies in the world that aren’t subsidizing and never subsidized you, not the sole one that subsidized you heavily before.

[Quinn Slack @sqs]

Tokens are expensive.

Most people don't realize how expensive they are.

Get good at turning them into something really valuable.

[Nicolai Baaring @nicolaibaaring]

After one day of unchanged use, it seems the new pricing model of @github Copilot in @code results in a 8-10x increase in costs.

Are you serious?

[Quinn Slack @sqs]

@RezMah Thanks for the report, sorry for the issue. Will look into it and fix the issues. In the meantime, you can make a new thread and mention this thread's ID or URL to keep working. (Also credited your account for the thread cost.)

[Quinn Slack @sqs]

@deftdawg Ah, sorry about that...will reply there

[Quinn Slack @sqs]

We had a ton of settings and secondary UIs to migrate to an (upcoming) new design for https://t.co/dO5fbw4sOy.

Amp made this so easy it didn't require any coordination or planning, just doing.

Everyone is used to when a product gets a facelift on the core screens but the settings screens stay ancient. That always sucked, but now with agents, it's intolerably bad and really makes you question whether to use the software at all.

[Quinn Slack @sqs]

@broady hah...it's like layers of sediment, the design gets older as you click deeper