Roberto Selbach

@tlakomy @thdxr I'd argue a person did do it. Claude didn't obtain credentials and connected to prod by itself.

@thdxr Then again, if a clanker could do this so easily, a person could do it as well, no?

we'll adjust to this new world but this is why i was saying the people responsible for security/safety are freaked out your company maybe lax now to let everyone experiment but those times are probably ending

@erd0xbc @sqs @AmpCode I don't believe 5.3 is available via API yet

@sqs @rslbch @AmpCode Is Deep now using 5.3 codex?

Hey @AmpCode, I think I'm liking using Deep mode better than smart now? I don't know if it got better for some reason or if Opus 4.6 feels like a downgrade in Smart 🤔 To be clear, this based on vibes only.

@thorstenball I have some friends working on it at keycard.ai

Someone please solve auth for agents.

@Pierre_Precieux @ylecun @elonmusk T'en fais pas, je suis canadien et je comprends très bien 😂

@ylecun @elonmusk Il faut être francais pour comprendre celle la 😂

💯

I wish the government put some money where its mouth is and make it a priority to build a Canadian SOTA model. We can't hide from AI, we should at least make it ours.

@josh_wingrove This guy needs new material. It's so repetitive.

yeah it’s back

Trump threatens major escalation against Canada (though it’s not clear what he means by a deal).

@lemire Fair enough. I do agree that doomerism (not only specific to AI) is quite popular in Canada in general and maybe QC in particular

@rslbch Mr Carney's folks have been in charge for a really long time and they totally missed the boat... Not that another party would necessarily have fared better. The issue is deeper.

If you live in Canada, not a week goes by without a new anti-AI rant. Though I started early in 2022 using AI in my introduction to programming course, we are now under a government moratorium in Quebec: AI projects have been generally suspended. Furthermore, my university recently posted its new recommendations for students: they should avoid using AI and prefer conventional tools. It is the general vibe among much of the Canadian elite. AI is bad. Let me be clear: all students will use AI for their homework. But if they want to be viewed as “good citizens,” they will deny having used AI as it is “bad.” Meanwhile money is flowing generously. The Canadian government probably cannot keep track of how many new endowed professorships it has created to pursue AI, nor can it keep track of all the grants and programs they offer to spur new businesses. Have your cake and eat it too. Hate AI but also become a world leader in AI. I see a parallel with various government programs to incentivize having children… in a culture that valorizes careers above everything else. The slope cannot be both negative and positive at the same time. I believe that these mixed signals go a long way toward explaining why a city like Montreal, that was largely viewed in the past decades as a hotspot of AI research, has produced little to no advancement in the middle of the AI revolution. Notice how early doomers like Elon Musk and Anthropic have now changed their tune. They don’t talk so much about AI destroying us… You can’t win a race while holding down the brakes. And, in Canada, the brakes are held down harder and harder with each passing day. Much has been said about the 'AI bubble'. American companies are going too fast, they will crash the market soon. Maybe. But it is easier to recover from a market crash than from a anti-innovation culture.

@thorstenball I had literally just figured it out after squinting like crazy. Thanks

@rslbch Engineering A Compiler (very good!) and Types And Programming Languages

@thorstenball sorry for the weird question, but I can't see books behind someone and not try to identify them. And I could not for the life of me find these two behind you 😂

Got them! Engineering a Compiler Types and Programming Languages(?)

@fluffypony @burkov I've got a similar experience. I wouldn't say Claude Code is useless, far from it. But my experience with Amp has been just amazing. So much better IME

I’ve tried doing real work on large codebases / working with a team on Claude Code. It’s beyond useless for both of those use-cases compared to Amp, which uses Opus 4.5 as its main coding model. Of course it costs us way more than a bunch of Max subscriptions, but we’re way more productive with Amp than we would be with CC. I fully believe that CC is a consumer coding product - that’s not a bad thing, and people will use it to create cool things, but once you’ve used a harness like Amp it’s really hard to go back to something like CC or Codex unless you’re financially obligated to (and then I’d just suggest using Amp Free, which is ad supported and gives you $300/month in Opus 4.5 usage for free).

I tested Claude Code with an API key where you pay for tokens. Every code update request for my apps costs about $0.80 in token cost. I make a couple hundred requests every day with my Max subscription ($100/month). If I paid for tokens, that would cost me around $80/day. So, Claude sells Max about 30 times cheaper than it sells Opus to third parties. There's no way a third-party IDE company without its own Claude Opus-like model can provide a competing agentic coding service to Anthropic. Cursor? Are you kidding me?

I decided to use @AmpCode exclusively this month. It's such a pleasure to work with. I absolutely love /handoff. But there's something magical asking the agent to do something and seeing it ask, unprompted, the librarian how something is done and seeing it go fetch a ton of websites.

@duailibe @sqs @steipete @AmpCode Fair enough. For reference, the Codex tab shows the balance of extra credits and a menu item to buy more (which just opens the browser on the OpenAI dashboard). I think for Amp it could look the same

@rslbch @sqs @steipete @AmpCode Yeah I had the balance but it was very different from the other providers (usage based in a window, replenishes at certain time) so I decided to split the PR to just show the Amp Free usage and them make more Amp specifics changes later

Just opened a PR to @steipete's CodexBar to add support for @AmpCode free tier github.com/steipete/Codex…

@sqs @duailibe @steipete @AmpCode Bonus internet points if we can also get the current balance 🤞

@duailibe @steipete @AmpCode Cool. We have an API for this but might remove it because it's not being used in the client right now. Just hit us up if you need anything.

@sqs @levifig Thats an inherently anticompetitive advantage for labs, isn't it? They can subsidize their subscriptions from (1) people who pay and don't use and, probably mostly, (2) API pricing users. Whereas others like Amp are at a disadvantage because at most they could have #1 🤔

@rslbch @levifig Yeah. If we end up supporting subs, then starting to warn people of our intent early and often will help, so thanks for asking and getting this out there.

There's simply no way Black is an alternative to Max 200 and Anthropic knows that… …BUT… …I don't think Anthropic realized that their biggest mistake was now having devs look at other models+subscriptions. Sure: Claude Max is "only" $200 (for $1000s at API-priced usage) but now devs can start looking at Gemini Ultra + CoPilot + ChatGPT Pro + Other Model Providers in OpenCode (and other harnesses). They will then start distributing their usage based on expertise, and that will mean Anthropic will inevitably lose mindshare and dev preference, which is something they had (and rightfully earned)! So, sure: I might now have to pay $500-600/mo instead of "just" $200, but I'll start getting better results, which may end up with me not considering Claude Max worth it, or Opus 4.5 worth it over other models! Even if the next Opus is excellent, the "pull" will be less dramatic… OpenCode, Clawdbot, and others (Amp should consider this too) should just keep pushing for their ability to use multiple models and model subscriptions. I fully believe the hockey stick-level curve will slow down (in terms of model quality), and cost-per-token will start coming down dramatically. The winner will not be the one that can charge the most for their models, but the one that has the "mindspace"… OpenAI has known this for a very long time (see their efforts in keeping up ChatGPT's popularity) and Google had that before in the search business… Interesting times ahead and I, for one, am excited!

@sqs @levifig Oh I see what you mean. Yeah, agreed. Then again, the biggest draw of Amp for me is how opinionated it is.

Yeah, let’s say we had Claude Max sub support and it was allowed. So you could use Amp by paying Anthropic $200/month and us (say) $40/month for the harness. Then let’s say OpenAI comes out with the best model (say GPT-6) for smart. Switching to GPT-6 would mean suddenly users pay full token consumption price, so their price would go up by a lot. And people would ask “can I keep using Amp with my Claude sub? I know it’s not the best model, but it’s a lot cheaper because I’m already paying $200/month to Anthropic for Claude Max.” We’d say no, and then we’d lose a lot of users, people who want a cheap harness MORE than they want (what we consider to be) the very best agent. We want to build exclusively for the latter group, those on the frontier. I’m not saying it’s a bad choice for other agents to make different choices. Just that it’s not the choice we want to make unless we can be assured of our flexibility to switch models.

@sqs @levifig Can you expand a little on "we might do it if it's allowed"? Maybe I'm confused by what "it" means in this context. Amp has changed the default model multiple times already, right? Why would this not be allowed? Again, I probably picked the wrong "it", hence me looking for clarif

Re: Amp, we might do it if it's allowed, but we would need to make sure that: 1. Everyone knows we're still free to change models when a better one comes out, even if the new model isn't covered by a particular user's sub and would be more expensive 2. We can still charge an Amp Token Fee to cover our costs (1) is the big one and is specific to Amp. This is not a priority now but we're definitely open to it.