Rob Smith

@itsolelehmann They won't, because making the product is a slither of the success story. Unless they start running targeted gtm and support for thousands or millions of niche applications they will remain the infra layer

TLDR: startups need to own their intelligence Why: you're currently paying OpenAI and Anthropic twice: once in tokens, then again in your proprietary knowledge they feed the models to make them useful The Fix: run open-weight models where nothing you type gets stored or trained on your prompts are a live feed of how your business works. your pricing, your workflows, your data, etc Satya says it plainly: providers reserve the right to learn from all of it so when they ship your workflow as a feature next year, you were the R&D department. and you paid them for it play that forward across every company and you get the bear case: OpenAI and Anthropic eat the whole economy, everyone else rounds to zero the good news is, open-weight models (intelligence you actually own, no labs sitting in the middle) are finally good enough to switch to. GLM-5.2 matches GPT-5.6 Luna on the Artificial Analysis index and beats Gemini (chart below) my personal favorite way to run these models today is on Venice. you pick the model, chats stay private, and nothing sits on a server to leak so you pay yourself back twice: once in saved token costs, then again in your data that never leaves your company

@p_millerd Keep calm and carry on

anyone have tips for hypersensitive 3 year olds having lots of tantrums and screaming?

@nateherk Wow. I have seen your videos and I could not tell on this example that this is stitched together with AI

One /goal prompt to GPT 5.6 Sol and I got this. I didn't script, research, record, or edit. I did nothing.

@maxkarpis - you probably saw, but sharing with you for visibility and reach within Revolut circles to get attention on this type of stuff. I am noticing Revolut missing things in production code lately that should never be missed. Come on @Revolut don't let standards slip now.

@GergelyOrosz Surely Google will release something big in the coming months to bring them back up with the leaders though?

Its not even a question that Google / Gemini is no longer in the top 3 coding model providers Which is surprising given the spend and resources Google has, and how they seemed to want to play to win Claude, GPT, and probably Grok and GLM seem to be the top coding LLMs

@trevin A lot of noise. I can't justify switching between anthropic and codex right now but I am interested in making GLM a part of my Claude code for the grunt work in anticipation of the higher token costs with fable

my feed right now feels like whiplash: "GPT-5.6 is the best thing ever" "Grok 4.5 is the bestest thing ever" "Anthropic models are the bestestest ever when you use Fable as advisor and Sonnet/Opus as orchestrator" "Look at our benchmarks of our own models that show we are the bestestestest"

@PawelHuryn Or just make Opus 4.8 executor with Fable as orchestrator. If max 20x limits allow it.

I doubt anyone wants to run Sonet as an orchestrator. A better way is to start with Opus and delegate down or escalate to Fable:

@GergelyOrosz Good analogy, or maybe a whole games console release every few months given how much we can do with one!

New AI model releases now remind me of AAA game releases: By launch day, all the game critics have been playing the game for weeks and publish reviews at the same time about their impressions. And tbf, these models cost as much (or more?) than AAA games...

@oikon48 My addition to claude.md/agents.md - "Verify delegated work before it counts. Check subagent and worker output against the primary source (open the file, run the test, launch the flow) before folding it in or reporting it done; a failed check sends the work back."

@oikon48 Yes just discovered this, and about to test it out. Have you found it works effectively? I saw some people complaining that it is lazy so I am experimenting with a claude.md prompt to ensure when work fans out to subagents or workflow workers, the coordinator owns the result.

もしかしてClaude Codeの /advisor コマンドで、指定のモデルをアドバイザーにできる機能が標準搭載されているの知られていない？？

@daniel_mac8 Thanks for sharing this - are you paying for the codex and claude code pro/max plans to facilitate this? I am on Claude code Max 20x currently, but needing to pay for the Codex pro on top is holding me back.

Try it here: github.com/DannyMac180/fa…

This is an amazing agentic engineering pattern. fable-advisor is a free, open-source Claude Code plugin that makes it even better. Empowers Fable to use: > Opus 4.8 > GPT-5.5 > Sonnet 5 ...as the implementer. Fable chooses based on task complexity. At the end, Fable does a read-only pass over the generated output for review.

But I do not intend to use Sonnet 5 as the executor. Opus 4.8 High is my preference.

This feels like the most important thing to implement in our setups before using any more Fable moving forward. I'll take 95% of the value for a fraction of the cost everyday...

Claude Code's cloud agents can't see what they build. No browser in the sandbox, so frontend fixes ship unverified. Copilot had browser + screenshots in cloud for a year. Cursor too. Feature request for Claude Code: github.com/anthropics/cla… @bcherny

@SahilBloom The last sentence is key. I believe we are better off keeping our stories to ourselves until and unless we have some proof. Proof breeds trust, and trust breeds opportunity.

The older I get, the more I realize how much of your life is shaped by the stories you tell yourself. I’m not ready. I’m too late. I’m bad at this. I never follow through. Be careful. You believe what you repeatedly rehearse. Tell better stories. Then prove them through action.

@github @impeccable_ai Thank you to @mvanhorn for inspiring me and giving me the confidence to speak up as a non coder when i see things that could work better. This was the interview i heard and i recommend anyone else on this journey to listen - youtube.com/watch?v=BxEf3R…

Big moment on my agentic engineering journey. I unlocked Pair Extraordinaire on @github for contributing to the @impeccable_ai open source project. AI democratises, non coders can now articulate value add ideas/improvements in an actionable format. github.com/users/makappen…

@msefaoruc i am out of town in Winchester but will give you a follow for any future events where i can get in.

who's building solo in london? thinking of organizing a casual build session at a cafe or coworking space. bring your laptop, work on your own thing, chat during breaks, and meet other builders. anyone interested?

@ryancarson For example - I still appreciate seeing my banking spend in a consistent way each month as a master view. Then can generate throwaway sub views where I want to dig in further.

@ryancarson I can imagine both co existing, it's not either or. There's value in a clean dashboard, if it isn't broken don't fix.. but now every edge case can get a nice view too. More software, more dashboards now it's so cheap and fast to create

Every time you think you need a dashboard to look at data, stop yourself. Do this instead: 1. Ask your agent to make sure that you have all the data to analyze something actually stored in the database. 2. Ask your agent to write a skill to gather that data. 3. Ask your agent to do the analysis and create a temp and throw-away HTML dashboard to answer the question(s) that you have In my experience, every dashboard that I've created gets less and less use over time and decays. It's much better to make sure your agent can get the data you need and answer the questions you have, on-demand.