message has been deleted

Great post @kbk_dk be interesting to see how the new haiku model is when it drops soon. I have found the same encoding issues and believe Phil's put fixes in for them..but might be worth raising on GitHub again. My agents ended up building a harness app around it to fix the issues and something else I had on Widnows but that got fixed

I was so surprised that haiku actually came back with a working website, but I was even more surprised that it then took so many more attempts to get a second one so the first one was just a super lucky take.. but yeah you can feel that I'm setting up haiku for a follow-up redemption arc blog post.

Brilliant!

Hey Kasper, Thanks for this blog post. It's excellent. For the small tweaks that you found could you add some issues. Also for the Haiku issues have you considered putting it into a loops where you tweak the mcp source and create a blog post until it finds some changes that make it work better for that model. I would be interested to see what extra it needed to get it right. Or the same with an Agent Skill to help guide the smaller agent.

Glad you liked it! I'll see if I can make some issues soon this week 🙂 As for Haiku, it was always just thrown in there for a laugh and to have something fun to do with Haiku 5.5 hopefully lands soon. I am sure I could make it perform better by tightening the rules, and tweak some stuff, but the point was never really that it should make it through. I know it's a bit content-farmy but I've also posted it on LinkedIn and would love a like or a comment in there - supposedly it will help the reach of the noise I'm supposed to be making for Umbraco DK Festival later this week (so sayeth our marketing person) Would be a big help and at the same time I feel a bit gross asking, so do not feel under any obligation 😅 https://www.linkedin.com/posts/kaspar-boel-kjeldsen-ab96992a_the-little-haiku-that-could-kill-every-share-7512794138392702976-lTY7/

I've replied and commented. Sorry if I added more ideas and work!

It's awesome, I signed up for a whole month of Codex for this so I'll have a few more weeks worth of Open AI tokens to burn before cancelling that sub again - Kraftvaerk is experimenting with sovereign AI and local models and I'm actually tomorrow being granted access to an in-office deepseek v4 flash that I'll be running through Claude Code, so I'm changing nothing until I've run that through as well. After that, we'll see 🙂

I would really like to see how this performs against a local model. Matt and my talk was accepted to the DK Festival. Look forward to chatting to you about it there!

Classic Phil move more work.

I'll be honest - it's mostly gonna be Claudes problem

https://discord.com/channels/1072196207201501266/1072197729276657744/1557453264067231784 It's alive and benchmarks well ofc

It won't surprise you it does well on the umbraco benchmark as well. I made a little-guys and a "little-guys-with-extra-handholdning" version to try to give haiku 4.5 a better chance as well as to try some open weight models, Haiku 5.5 needs none of that and does quite well on the regular mcp test.

Awesome look forward to the results 🙂

Currently doing some extra-handholdning "Devstral 2 2512" which is ancient in model-terms now, but doing suprisingly well when it gets good instructions

Impressive what Haiku is capable of. Imagine what the next one will do...

I was honestly more surprised by huge gap between even Haiku 5.5 and basically any of the open weight models. Cutting Mistral 4 big some slack because it was launch week and the connectiong/latency was acting weird and got rate limited in some of the runs, but none of them could reliably do what Haiku 5.5 did without a lot more help. Now, they all ran through Claude Code harness, so it's not entirely fair, and I'm open to trying some other combinations at some point.

Yeah, it feels like one of the bigger leaps from a previous version of a model as well. Seeing the open weight models been so behind makes you wonder how much of the power is distilled

Muse Wonders did not do amazing.