Henning Naarlien-Tolpinrud

Data, AI and retro.

[Henning Naarlien-Tolpinrud @durniplot]

@PRingholm Nå som veldig mange modeller kan levere «bra nok» AI er det hvor dataene bor og man jobber som teller. Post fra starten av juni. https://t.co/jvXqJODI0G

[Henning Naarlien-Tolpinrud @durniplot]

@PRingholm Vi snakker kanskje lite om det fordi endringen i stor grad skjer langt unna oss? Vil tro de er mye mer klar over dette i Kina, India og sør-øst Asia. Landene med lavest levestandard er også preget kriger som overskygger det meste annet.

[Henning Naarlien-Tolpinrud @durniplot]

v0.9.3 of the @cline plugin adds *experimental* @OpenAI Codex support, with some extra steps when setting up. In my experience, Claude is much better at delegation and orchestrating and not stopping, but I might just have skills issues. Been a while since I used the codex cli.

[Henning Naarlien-Tolpinrud @durniplot]

v0.9.0 gains concurrency improvements and tighter guidance for the orchestrating model. With all the delegation/auto-routing buzz, it feels like we are close to good enough until this is just built in everywhere.

[Henning Naarlien-Tolpinrud @durniplot]

@dr_cintas You can do the same with Cline and access cheap ClinePass inference (or whatever else Cline supports) with https://t.co/MLH3402I7X and there is also a skill for delegating til opencode out there (for Opencode Go, etc.). Have Fable/Opus route tasks by required skill+cost.

[Henning Naarlien-Tolpinrud @durniplot]

v0.9.0 gains concurrency improvements and tighter guidance for the orchestrating model. With all the delegation/auto-routing buzz, it feels like we are close to good enough until this is just built in everywhere.

[Henning Naarlien-Tolpinrud @durniplot]

v0.8.0 adds support for generating profiles for e.g free models on openrouter and others, so you can spend even less credits. You will need to serve your own feed. Spec and code you can start with is here. https://t.co/OyLLX0yAjC

[Henning Naarlien-Tolpinrud @durniplot]

@Yuchenj_UW You are not wrong, manual model routing is just too much of a bother, but codex and claude are pretty good at delegating tasks if you give them enough info to do it (like you also do with Polly). I ride Claude, which delegates to Codex and ClinePass. https://t.co/ePdKQmxMJJ

[Henning Naarlien-Tolpinrud @durniplot]

@theallinpod This podcast has the potensial to be really great if they found some knowledgeable researchers and hosts that weren’t just discovering weeks/months old trend now.

[Henning Naarlien-Tolpinrud @durniplot]

@cjzafir Why use the smartest model as the orchestrator? I use Sonnet for orchestrating the builds, after Fable did the PRD, plan and task breakdown + matching model to task. Sonnet just farms out tasks to the various claude models, and codex and cline plugins, and uses opus/fable for QA.

[Henning Naarlien-Tolpinrud @durniplot]

v0.8.0 adds support for generating profiles for e.g free models on openrouter and others, so you can spend even less credits. You will need to serve your own feed. Spec and code you can start with is here. https://t.co/OyLLX0yAjC

[Henning Naarlien-Tolpinrud @durniplot]

Now updated with quite a few improvements. Ran out of @cline pass credits for the week while building though. 🤷‍♂️

[Henning Naarlien-Tolpinrud @durniplot]

@bios_hazard @mattpocockuk Absolutely. Agents are what let the new software factories to build bespoke software almost autonomously, as opposed to the old ones that used lots of people to customise standard parts or build on top of a broad base like Salesforce or SAP.

[Matt Pocock @mattpocockuk]

Four of AI's most confusing terms explained:

An agent is just a model, harnessed, in an environment.

[Henning Naarlien-Tolpinrud @durniplot]

@bios_hazard @mattpocockuk I wrote a little about software factories here. https://t.co/6Dv0HLrMEm

[Henning Naarlien-Tolpinrud @durniplot]

@VagKos @mattpocockuk This might help. https://t.co/uxQ2ivH4Tu

[Henning Naarlien-Tolpinrud @durniplot]

@cjzafir Planning with Fable, then orchestrating with Sonnet, using Codex and Cline w/Clinepass and reserving Codex, Opus and Fable for reviews and escalations gets you even further. https://t.co/MLH3402I7X

[Henning Naarlien-Tolpinrud @durniplot]

@0xStacking @8teAPi Currently my workflow is to use Fable for spec and orchestration planning, then I switch to Sonnet for execution with instructions to use delegation and to match tasks to the cheapest model it thinks will work, and reserve opus and fable for reviews and emergencies.

[Henning Naarlien-Tolpinrud @durniplot]

@0xStacking @8teAPi It has worked well for me. Speed is OK, output quality had been OK, it gets the job done. I don’t need neck breaking speed I am just a pleb building ordinary stuff. It is good enough for me and makes my Claude and Codex go a lot further. Friction is using a new (to me) harness.

[Henning Naarlien-Tolpinrud @durniplot]

Now updated with quite a few improvements. Ran out of @cline pass credits for the week while building though. 🤷‍♂️

[Henning Naarlien-Tolpinrud @durniplot]

With a fresh @cline pass in hand, inspired by the Codex cc plugin, I had Claude and Codex w/@shadcn and @mattpocockuk 's amazing skills whip up a Cline plugin so Claude can delegate tasks to it, and save on Claude tokens. Works for me, maybe for you too! https://t.co/MLH3402I7X

[Henning Naarlien-Tolpinrud @durniplot]

With a fresh @cline pass in hand, inspired by the Codex cc plugin, I had Claude and Codex w/@shadcn and @mattpocockuk 's amazing skills whip up a Cline plugin so Claude can delegate tasks to it, and save on Claude tokens. Works for me, maybe for you too! https://t.co/MLH3402I7X

[Henning Naarlien-Tolpinrud @durniplot]

@omnigent_ai "Intelligent model router to pick the right model per turn" is missing from the release notes on github - did that get cut for now?