alfalama

[alfalama @blekachu]

@DavidMinovsky I checked and It's about 10% in the last 7 days so it didn't make the cut the top 6 listed in the graph

[alfalama @blekachu]

basically switched from only using gpt models to claude

[alfalama @blekachu]

SGlang and Nvidia meetup in San Francisco tonight

[alfalama @blekachu]

My grok bot was forgetting stuff when I threw more at him so I have these open loops that he keeps track of and for the past 2 weeks its been working great. Try it out if yours forgets stuff:

Everything unfinished is one row in one file. Done means a documented result, not "I read it" or "I prepared a draft."

The register is ~/ops/loops.csv on the bot's computer. Each row has id, area, outcome, and owner; state (working, waiting_on_me, waiting_on_them, blocked, done, cancelled); next_action, urgency, due_at, and next_check_at; source_refs (the mail, chat, or thread it came from); plus draft_ref, approval_ref, completion_evidence, and updated_at.

Only the main bot writes. Other bots hand off the event. Writes go through a small Python script with a file lock so the CSV does not get corrupted.

A loop can be closed only with completion_evidence (for example "mail sent, message id …" or "payment received"), or when I explicitly cancel it. A read email, a status report, or a finished draft is not done. The loop stays open until the thing actually happens.

When something moves (a sent mail, a reply, a booked call), the bot updates the loop in the same step. It does not wait to be reminded. The bot sends nothing on its own. It prepares a draft and I approve the exact text.

Routines that read the register:
- Daily action list at the start of the day: what is waiting on me, what is overdue, and what to check today.
- Two approval packs a day (afternoon and evening): decisions and follow-up drafts with last-message context. I just write the numbers to send.
- Immediate ping for a cancelled call, a client problem, or an unanswered reply to a cold email.
- One-off reminders: "ping me on Monday to check X" and the bot creates a routine for that day.

[alfalama @blekachu]

my macbook is usually overloaded from all the agents running so I'm trying @boatdotdev today for typecheck, E2E and builds.
Stack will be T3 code + boat for builds and tests.

[alfalama @blekachu]

@vineerpasam Thanks, I ran the prompt and had agent explain it to me but your description was way better

[alfalama @blekachu]

@_dylanga Thanks man, improved membership access round-trips to DB
22 files / +365 / −274

[Kruspin @Kruspinapp]

What we shipped at Kruspin in the last two weeks: Buxus, an AI assistant. We also added a Canva integration, QR codes that let guests find their own table, and new seating plan elements like a stage, bar, tent and ceremony arch.
Sign up at https://t.co/26YZyUWcw0 and try the demo

[spoony @spoonyvu]

I have hundreds of photos of my daughter i love and almost never share.

I wanted to share the moments, not her face. but the emoji-over-the-face thing always felt silly to me and kind of ruined the photo.

so I designed and built Peekaboo. It gives faces a little disguise instead, like sunglasses, a cap, or her own hands.

the iOS app, live today 👇

[alfalama @blekachu]

Lost my sales rep for Kruspin in Czech Republic. Not sure if I should try to
- hire another one and find 2-4 more in other countries
- go all in on AI and take care of the sales myself for as long as possible

I'm leaning towards the second option but don't want to do the job half assed

[alfalama @blekachu]

@poteto trying it out for the first time and gmail connector doesnt allow sending emails From as aliases that are added in gmail.

[Robert Lukoszko @Karmedge]

i found perfect instructions to my AI agents :

[alfalama @blekachu]

@Howaboua What do you think about the agents api? I'm trying to wrap my head around where to start with building AI into my app

[Howaboua @Howaboua]

Right, I have now slept, am on my 2nd coffee, so I can actually share my thoughts on devday, for those interested.

No, it wasn't crap. But what they are launching was definitely rushed - both in functionality and in comms (see the tweet when they said that Astra allowed them to push things 6-8 months ahead of their planned launches).

Dots

The free-ish usage on Dots for the first month is to collect data on what's working and what isn't and what the user patterns are for using these, and how much their capacity's gonna hurt, before they decide how to charge for them and how to route models, and if they even have the capacity to do more dots per user. I don't mind/care, as misleading as it was. The problem is, we're yet again beta (if not alpha) testers. Dots will look VERY differently (quota-wise) once they don't need to gather user feedback anymore. I have my doubts they will run on Astra.

I don't even have access so I can't say whether it's good as it is or not. I'm sure they will iron out the kinks. I think this is targeted towards normies locked into GPT ecosystem. Non-normies already have their alternative. It might be less functional, but it's tailored to what they need. The only temptation comes from not drawing from the quota.

Personally, I'd only be using it with Discord, WhatsApp, and maybe Telegram if they force me to. I don't have access to Muse yet, from the demos I've seen, I'd rather have Muse on my phone (assuming Android integration is as tight as iOS), and just let it use my phone as if it was its own (ish). So if anything mentioned in this para isn't there for dots, Muse is the better choice. I am also bullish on Muse Spark being truly multimodal.

One thing that worries me is that the dot will have access to every plugin in the ecosystem if it looks it up. Which means 95% of them will not get used. And a lot of them will not be quite as optimised as they could be. There's also the problem of functionality overlap. Like idk, if you're using Tidal and not Spotify, and dot suggests Spotify once, it should now be blacklisted until further notice from the plugin search, if you catch my drift.

Also: just Tailscale into Herdr or run Herdr locally, forehead. If you're paing for a gippity sub but don't like Codex, this is the way. I feel like this will be the poweruser's dream and I will be on board with this. If they restrict it... Welp. I'll scream and shout.

Sol 6.1

Too early to say. All things are pointing towards it being the Astra Minor we've seen. Sol 6 seems like it was closer to 5 family in everything, just RLed/distilled by a bigger model. 6.1 seems to be closer to Astra in outputs so far.

Doesn't change much in my workflow. Since Astra dropped, I've been running it as the main agent and finally allowed subagents in Pi (via herdr tabs), because that model knows what it's doing with them.

I've been using Sol for implementation with it, 5.6, then 6, and now will be using 6.1. I think I am finally willing to encourage using implementation subagents more and see where it goes. For pair-programming sessions, where you go back and forth, Sol 6.1 is going to be a banger and if you're using short, scoped prompts (early 2026 style), this is your main model, don't even worry about Astra.

And cache is becoming more important than ever, since that's where the savings come from, so if you're running a 3rd party harness, make sure it's as close to optimal implementation as poss, and that your extensions are not nuking the cache with every api request.

Quotas & limits

Looks bad on the surface, but let's just all pour one for the subsidised subscription meta. This had to end, just like request-based subscriptions. The earlier this bandaid is ripped off, the better for everyone. Users living their unlimited tokens delusion, labs having to constantly struggle with compute... Imagine a world where the subsidies continue endlessly. This is what is fueling the bubble. This is what one day can get us into a situation where running an AI lab isn't profitable anymore and a frontier lab just shuts down because nobody wants to invest in it anymore. This is what could lead to sovereign labs getting bought out by governments and nationalised (like, truly nationalised).

Yup it sucks. We kinda have to deal with it if we want the AI revolution to continue. I appreciate the commitment to making the models so efficient they subs just won't be a thing anymore. Meta could do it with their contributor tier. I bet OAI and Anthropic could introduce one too, but the implications of ppl not understanding they're fully agreeing to sharing stuff with them are too big imho. Let's just hold OAI to the efficiency promises - they've been doing quite well so far.

Plugin extensions, sites with plugins, ChatGPT space, etc.

Very good move. This is potentially what could bring 3rd party GUI interface people into Codex. Massive unlock for companies and people who don't actually use AI to code. I am a big proponent of " AI Coding Agent Software" not being dev-oriented. Yeah I know it reads weird, but if you are an accountant you don't need/want to know what git diff or a PR is - you want your spreadsheets and custom dashboards to be displayed next to your chat window. This is the first step from a major company. This ties nicely into...

Proper Sign in with your GPT sub

ABSOLUTELY LOVE THIS. SaaSes that charge you premium for their AI tier that uses the shittiest model available need to disappear. I've been saying it for a while now - if your SaaS doesn't allow BYOK with your own model selection, you need to find another one ASAP. This is where issues happen, like you don't want to be forced to have Gemini Flash Lite operating your CRM, writing outreach emails etc, just because they found it to be the most economical model. And don't get me started on the possibility of the SaaS not upgrading to latest, because it's SO HARD TO REPLACE A STRING from "3" to "3.1". PPL have been saying that SaaS is dead - but maybe this is what will revive them. Because no, vibecoding your own alternative to a managed service is not really a good idea at scale.

Ultrafast - not for me. It's interesting that ppl are saying it's not running on Cerebras.

Decisions API - sure. I will have more thoughts when we find out if we can use it via our GPT sub somehow for stuff (unsure if you realise, but you can just call your GPT sub as if it was an API if you try hard enough)

Private intelligence - I wonder if this is gonna unlock using GPT for some European companies that simply are not allowed to use American or Chinese models (first hand info). We shall see.

Codex in the cloud (v2)- nice (esp with dots), if you are using Codex with your GPT sub. I'm the biggest AMP shill that doesn't even use AMP. And AMP seems like the better solution for Cloud tasks, esp for teams. So far there are no signs that it uses up more quota than running Codex locally (as opposed to v1) - a big win if this holds up.

Finally, the most important question - should you cancel your sub?

I don't fucking know, you do you. I still have a lot of my resets (and now credits) to go through.

The whole problem with the limits is that you are probably using Codex, which isn't as efficient token-wise as it could be. Yeah it's feature-packed and it can do anything, but just how much of that functionality are you actually using? The 5% can be ported into Pi or you can create your own custom harness, or tweak something small like FX, or build your own GUI. You're reading a post from a guy who's done it. It's not that big of a problem if you know what you're doing and it seems like ALL of the X bubble are experts on everything harness and coding related, so just knock yourself out and do your own thing.

If you can live with Claude Code, it's probably a better sub right now (jury is still out as far as Sol 6.1 goes, quality/quota consumption-wise). Might not be in a month. Or it might be. Problem for me is, the minute someone locks me into their software, their toolkit I'm out. I want frictionless interactions tailored specifically for me. Nobody likes to be forced into a specific ways of working, have you ever worked a job in your life? :P Why would we do this with a coding harness, a tool. We're already have that friction issue everywhere else.

The Meta contributor tier API is staring in my face though and is looking mighty tempting. Muse Spark might not be the best model for coding, but for talking to and personal assistant stuff, it always wins. I've tried my discord bot on GPT models. They all sucked and sounded robotic. So that's something to consider if you plan on using a dot heavily.

Idk, I could be rambling more, but ain't nobody gonna read this anyway. Go away now.

PS.

No, I am not an OAI apologist, I'm just realistic, unlike some fucking hypebros. Some of the things we are disappointed by, we had coming, we just didn't want to face the reality that shit is never free in life. So a "deal with it" attitude might not be what we want, but is what we get from life. And we either deal with it by simply accepting it, or... finding alternatives. However you approach this is up to you to decide. No company is your friend, ever. We live in capitalism era.

I am trying to NOT farm hate/likes/whatever here. I've got a sub with GPT because it offers SOTA models for 3rd party harnesses and allows to plug it in anywhere I want, and I will die on the hill of defending this openness.

[Isaiah Granet @zaygranet]

A few thoughts on switching lines. https://t.co/3JIXR2G4eJ

[Rob Miles @robertskmiles]

If you're going to send me LLM generated text, please just send me the prompt you used to generate it instead.

This isn't even a joke, I would genuinely be happy to read the three bullet points you turned into that whole email

[alfalama @blekachu]

@synopsi Red version for the win

[alfalama @blekachu]

@synopsi yeah maybe some little breathing between. I meant more the pauses than the voice itself.