polymorpheus

@gilgNYC when do you think we'll see something big and frontier-like from overseas, e.g., a new GLM, Kimi, DeepSeek, MiniMax, Qwen

My thoughts: Anthropic is preparing Opus 5 and need to release it before removing Fable 5 from subs to compete with GPT 5.6. And so we’re probably looking at an Opus 5 release within the next two weeks.

@AniC_dev gonna have to give 📦 a spin!

easiest is timeout after agent round but there are as many ways to do is as there are usecases and builders the smarter you are with it and the more tailored to your usecase the more you'll save costs it's your machine, you get sudo, you could even spy on processes to know when is best time

maybe you don't always need it, but the best is to do it tho, for convenience why plug a lot of tools manually when you can just run the agent on a computer that has them all, and where the agent can improvize new ones? models are getting smarter and more autonomous than us humans trying to craft MCPs and stuff free the agents! also if you think that's expensive waste, it's because other providers overprice a lot box is 5x to 25x cheaper at $0.00001/s if you strategically stop/resume the sandbox in between agent runs, you can spend < $1/user/month

@AniC_dev there's definitely an interesting use case here for, if not several. e.g., spin up a box, add the repos/tools i need (dotfiles?), add sol 5.6 ultra and fable (is anthropic ok w/ that? can never tell), set a /goal, work while i'm sleeping!

@AniC_dev > if you strategically stop/resume the sandbox in between agent runs, you can spend < $1/user/month can my agent strategically stop/resume the box for me? is that just a cron? 🤔

probably closer to Spring 2025 tbh #Models" target="_blank" rel="nofollow noopener">en.wikipedia.org/wiki/Claude_(A…

@devinjameson might interest @SandroMaglione

State machines just landed in foldkit/experimental. import { Machine } from 'foldkit/experimental' Foldkit's architecture (TEA) keeps updates clear, but complex transition logic for ADTs gets gnarly and hard to trace. You need to match on the Message tag, match on the ADT state, guard appropriately, and repeat for each arm. Machines define every transition between ADT states in one place.

@RhysSullivan *tries to click email, clicks ad instead

shipping ads into my gmail that cause layout shift is crazy work

our team's first week of having both sol and fable the crazy thing is 30% of our cost was on fable

@captainsafia Feels significantly harder. Especially in a team. There’s like 2x the volume of code to review coming at 10x the pace. Code review isn’t possible or celebrated anymore, it’s mostly ceremony.

Vibe check for the room: are we all actually feeling like the work of maintaining software has gotten easier? We’ve got a plethora of harnesses, super robust models, all sorts of form factors and integrations…is it all clicking into place?

@samgoodwin89 > emotional energy on convincing people to do something (and failing) too real... > I ignore most of what's written and quickly find what I agree or disagree with yep - there's no other way - exhaustion, fatigue, and burnout are the alternatives...

We're all being forced to become directors. People worry that AI will atrophy the brain, but I think it will actually level us all up. I am just now realizing how terrible I used to be at explaining to others what to do, how and why. My nature is that of a lone wolf coder. It was always easier for me to build it than to describe it. AI has changed me drastically, and for the better. First off - I'm no longer afraid to be direct, precise and commanding. I've been told many times in my career that I am "too blunt", which I don't totally deny, but it's profound to see how my development process has evolved now there's no longer a human on the other side of my "blunt" instructions. There's also no longer any skill issues - I know that if I can describe it properly, then the agent will be able to do it for me. So now I really care about explaining myself in writing and have noticed my thinking change - I dream up agentic processes I can run instead of code I can write. I wouldn't put this level of effort in at a corporation because I didn't believe it would make a difference. I was never influential enough to have that kind of pull and couldn't be bothered trying to play politics and steer an org. I just wanted to build. Now that I don't have to spend all that emotional energy on convincing people to do something (and failing), I have so much more bandwidth to architect processes that produce the outcomes I want. Just like a director might do in a typical corporation, I spin up teams of agents to retrieve and synthesize information (code, docs, articles, etc.) into what I need (designs, plans and prototypes), which I then feed into implementation (code, tests, docs, the rest). It used to piss me off when the leadership at Amazon would give you like 5 minutes of their time and never try to really understand what you're doing. But now I totally get it lol. It's just impossible to operate at that level and care about every tiny detail. When you start orchestrating teams of agents, the amount of information quickly overloads you and forces you to filter. Agents produce amazing, but also long and extremely detailed reports. When reviewing this mass of information, I am forced to zero in on what is critical and prune as much of the details as possible. I ignore most of what's written and quickly find what I agree or disagree with. Just like my director would do to my over-eager former SDE1 self. I focus on the shape and direction of the solution. Quickly get a feel for whether it aligns or misaligns with my vision and go back and forth until it's good enough to try. Each iteration is even more information to review and understand. So again, you can't read it all. Once I am happy with the proposed solution, I then lay out how I want it executed. I like to describe workflows with steps and iterations - e.g."use sub agents for XYZ", "do this, then that", "review the outputs and produce a final ABC", "then come back to me", etc. Kick them off and then just monitor to ensure they follow good software development practices. Review the critical pieces of the code, make sure it's not utterly insane, is well tested, uses the coding conventions already established in the repo, etc. And if it's garbage after all that, just throw it away or feed it into a new idea. Code is fluid now and there's no one to be butt hurt about wasting 3 months of their time on a shower thought.

If you set gpt-5.6-sol to "ultra", all the subagents it spawns will also be set to ultra. IMO this is a fumble. Causes massive token burn for no good reason. At the very least, I should be able to hard-set the subagent effort level to "medium". Claude Code is far ahead here

@samgoodwin89 we need this for effect-smol lol

Gives me 87/100 and then brutalized me.

@dsgnevrywhr The main thing I can take away from this is that I'm not trying enough shit

@blackboxai - i bought a $40 account - how do i use this thing? conductor? opencode?

@ericclemmons Can you share a screenshot?

I admit, Effect didn't "click" for me (for the usual knee-jerk reasons). But, v2.alchemy.run showed me what a clean, composable API could be with it. With effect-smol cloned, I was able to wire Drizzle + Effect into a Hono app. Anytime I was confused, I just asked my agent "explain this" or "is there an Effect way of making it look like ". The result is actually *very* understandable!

@maxedapps how does grok 4.5 compare to… - gpt 5.5 medium? - opus 4.8 high?

Did more testing. I'm honestly impressed by Grok 4.5. Definitely was not on my list of models to be excited about this week. Well, I guess only GPT 5.6 Sol was on that list... And Fable, of course. Anyways, Grok 4.5 is really, really good! At implementing (it's fast) but also does a good job at reviews and planning. I'm bouncing back and forth between GPT 5.5 (I'm a peasant, don't have 5.6 yet), Fable 5 (for the more complex tasks and plans), and Grok 4.5. Much, much fun!

@BHolmesDev what about bandwidth + limits + cost? can i use it for more than 2hrs without running out and exhausting my budget/limit?

Been using Grok 4.5 for the night. It... genuinely feels like Opus 4.X, but with token output speeds closer to GPT 5.5. I'm super impressed. I'll throw it at more complex coding tasks tomorrow

This model looks impressive on paper. However, Meta has no history in dev tools, save for shutting down tools devs loved and used (Parse, 2013) It will take more than a single release to gain developer trust, and love. It takes years. I wonder if Meta will persist that long?

@LonePasserby @mattpocockuk Well said! Echoes the “strategic versus tactical” dichotomy that folks sometimes bring up

I think on this topic, people will have vastly different opinions whether they are typically doing work solo vs working in a team. Also factually speaking, to get a PR on par with a really good human-made PR, you’d have to have a dedicated PR agent who is fed 1) the overall vision of the projet, 2) the whole git commit history, 3) minute details from the designer about the work increment (coming from grilling, …) 4) insights about what is more “important” or impactful for the team in terms of communication. This last part is what humans do best and LLMs struggle with. They can’t “read the room”. But yeah most of this can be definitely improved, but most people don’t have a dedicated PR description agent eg

This is crazy to me AI is an incredible lever for making PR's and commits easy to review You just need a bit of steering to customize it to how you want

@RhysSullivan @zeeg Cramer (Sentry) shared something recently about training models exclusively on coding (no general-purpose) and it got me tinking: what if we had a small Effect model, like a gemma-effect-a4b? Then I could write Effect on planes and trains.

@zeeg now that the LLMs are writing my code i like effect a lot more, just give it some good references like the opencode repo, executor repo, and effect-smol source and you should get good outputs without too much pain

should I rewrite my project in Effect or will i just get annoyed