Most software work starts in Slack. A bug report. A customer request. A “can someone look at this?” Then the context gets scattered, ownership gets unclear, and someone has to recreate the task somewhere else. Alan’s Slackbot fixes that. Turn a Slack conversation into an Alan task with the context attached, assign ownership, and send it straight into the software factory: Plan → Build → Review → Test → Ship No copy-pasting. No lost context. No “who’s taking this?” Work starts where the conversation starts.

Been playing with Dot and I think there’s a really interesting product hiding here. The good: > it actually feels like an assistant rather than another chatbot > having its own computer is wild > the “call your Dot” UX is ridiculously well thought out > it proactively picks up work instead of waiting for another prompt > the attention to tiny details is probably the best part But the rough edges are very real: > painfully slow for simple tasks right now > computer use can take absurdly long routes to accomplish basic things > the cloud computer/permissions model still feels limiting > I really miss having multiple agents running in parallel > some basic UX around approvals and knowing what it’s actually doing needs work PS: I just told Dot to go through my Agentic Engineering YouTube channel (go subscribe plz) and figure out why we’re not getting enough views. It opened the channel, looked through the videos, thumbnails and analytics, and is now finding things we can actually change. This is the kind of AI use case that makes me excited for the future!

demn @sama gave Dot its own desktop and it looks so damn beautiful. the tiny details, the wallpapers, icons, animations, even the little Dot touches everywhere the level of thought that went into this UI by @OpenAI is crazy!

Mark this podcast as essential if starting an AI based Startup We sat down with @prasad_pilla , Founder & CEO of Alan AI and Co-founder & CTO of SupaTest AI, for a deep dive into his journey through technology, Gojek, and the world of AI startups. From his early fascination with technology to building AI-powered companies A conversation packed with insights for anyone building, exploring, or investing their time in AI,Prompt Caching,SDLC,ADLC, Orchestration,Open models etc. Watch the full conversation here: piped.video/PN4CuFhRmng #AI #ArtificialIntelligence #Startups #AIAgents #TechPodcast #Entrepreneurship #AlanAI #Gojek

$25 vs $0.60 per million tokens. same league on benchmarks. open models just got very hard to ignore.

RESET @thsottiaux

this is cool

you can now bring your ChatGPT plan into other apps. Devin running GPT-6 Sol on your existing ChatGPT plan is kinda wild. the model is becoming the commodity. the interface you choose to use it through is becoming the interesting part.

Dario says we should pace the frontier. Anthropic, approximately 6 business days later: > Opus 5.5 Sept 22 > Sonnet 5.5 Sept 28 > Haiku 5.5 coming next. To be fair, “pace” is apparently doing a lot of work here 😂 Anthropic launched Opus 5.5 on Sept 22, then Sonnet 5.5 on Sept 28. Sonnet is 30%+ faster than Sonnet 5 and up to 30% cheaper for most workloads. Anthropic says Haiku 5.5 is coming in the next few weeks. And today OpenAI has DevDay, so naturally everyone is waiting to see what gets thrown into the arena next. There are definitely multiple announcements/rumors floating around, but the official DevDay page is still keeping the details broad. The funniest part of the AI industry right now is that everyone agrees we should slow down…and then someone ships a better model before lunch. But underneath the memes, something important is happening: Model cycles are getting absurdly short. Opus 5 → Opus 5.5. Sonnet 5 → Sonnet 5.5. GPT-6 Astra → Sol/Luna → whatever comes next. If your product strategy depends on one model staying ahead for even six months, good luck. The real game is increasingly becoming: build above the models, not around one of them. Models will keep changing. The orchestration, context, verification and workflows around them are where things get interesting. Welcome to “pacing the frontier” 😭

building has become insanely easy. so everyone says distribution is the MOAT now. but i think even that has an expiration date. today, growth = get humans to discover you → grab their attention → convince them to buy. but imagine my personal agent looking for a dev tool. it doesn’t care about your viral launch, 50k impressions or a Product Hunt badge. it talks to your agent, compares you with 20 alternatives, checks pricing, benchmarks, security, negotiates a deal and picks whatever fits me best. now the game changes from: “how do i get humans to choose us?” to: “how do i get agents to choose us for humans?” growth won’t disappear. the unit of distribution changes.

AI is moving faster than most engineering teams can reorganize around it. Had a great conversation with @Aarav_network about what changes when copilots become agents. We covered: > why prompt engineering is becoming harness + context engineering > how we run 20-30 agents across engineering work > open models, subagents, context compression & AI cost > the shift from SDLC → agent driven SDLC > why teams may need an orchestration layer above Claude Code, Codex and other agents And one idea I keep coming back to: Software engineering was never about typing code. It was always about solving problems. Watch the full podcast 👇

OpenAI is rumored to drop its new agent, "o," tomorrow. It already has 3 competitors. One of them just raised $1B today. The agent war has officially started, and you're probably still figuring out which side to bet on. They all promise the same thing, but they're making very different bets. A month ago, Instinct announced a round that valued the company at $2.5B. Today, they raised another $1B at a $10B valuation. It has no UI. You just text it, and it gets work done for you. It runs in the background and reminds you when something important needs your attention. Meta is taking a different approach with Muse — a consumer app that even comes with a Tamagotchi-like keychain. SpaceXAI's Grok Bot is playing a different game altogether, acting more like an orchestrator for other agents. I'm personally betting on Grok Bot. Who are you using, and why?

Every time agents get more capable, the same curve shows up.

the opus 5.5 week, in order: Day 1: holy shit, this is good. Day 2: this might actually be better than Astra. Day 3: okay… it’s doing a lot. Day 4: wait, why did it touch that? Day 5: what the hell is it doing? they probably didn't nerf it lol in week one you give it demos. there's nothing to get wrong, so when it takes initiative, that's the magic. then you point it at your real repo, and the same initiative turns into guessing. it doesn't know why that module is weird, what broke last time, or what you meant by "just fix it." so it fills the gaps with confident assumptions and runs. the model didn't get worse. the tasks got real, and the missing context started to show. a more capable agent needs more context and clearer checkpoints, not less. that's what we're building at Alan AI Full article below 👇

it's just me and @tryAalanAI against the whole world atp

Codex reset just came in. Weekend sorted 🎉

there's a very fine line bwtween vibe coding and agentic engineering

the opus 5.5 week, in order: day 1: "5.5 > 6" day 3: "it's doing too much" day 5: "i didn't ask for that" day 7: "they nerfed it" they probably didn't nerf it lol in week one you give it demos. there's nothing to get wrong, so when it takes initiative, that's the magic. then you point it at your real repo, and the same initiative turns into guessing. it doesn't know why that module is weird, what broke last time, or what you meant by "just fix it." so it fills the gaps with confident assumptions and runs. the model didn't get worse. the tasks got real, and the missing context started to show. a more capable agent needs more context and clearer checkpoints, not less. that's what we're building at @tryAlanAI

i still can't wrap my head around this just 1 year ago, I paid $1,000+ for a video like this now I made this with Opus 5.5 > in less than 30 minutes > costed about ~$24 (fraction of what i paid) this is just BIZARRE!

motion designers rn

this might actually explain why Opus 5.5 feels so ridiculously good at motion design. it’s not just generating pretty visuals it can reason through the entire process: composition, timing, typography, movement, visual hierarchy, code, iteration, etc. the CoBench results are a pretty good signal of how strong it is at complex, multi step work. and that’s exactly what motion design needs. we’re still not at “replace a professional motion designer” territory, but the gap is getting very uncomfortable. 😭