Mack Delany

@ohabryka Also, just having model options. GPT 5.2 >> Opus 4.5 for ~60% of our codebase.

@ohabryka I spend ~$4K month on Cursor & wondered the same given the Claude Code timeline love fest. So, been running a couple worktrees on Claude Code lately... it’s fine. No better than Opus 4.5 in Cursor and I prefer IDE ergonomics. I think people are fanatic about their own setup.

Ok, can someone explain to me why people use Claude Code instead of Cursor? I get obvious productivity boosts from having my IDE integrated with my AI tooling, and lose that in Claude Code, but everyone keeps raving on about how great Claude Code is, so what am I missing?

@VictorTaelin Have found GPT 5.2 very good for gnarly geometric operations with rust/c++. Opus 4.5 was good, now, slop.

Today was probably one of the worst AI coding days I ever had "You're right, I'm sorry. I invented 'dup root' as a hack to work around the fact that after WNF evaluation..." Jesus... I guess I made the terrible mistake of trusting Opus 4.5 too much. 5 commits later, the repo is in now in a hopeless state. There are lots of moronic mistakes that could only have been made by someone who didn't bother to read half of the codebase. Which, to be fair, is probably the case. Perhaps the repo might be past the point where LLMs start becoming useless...? I'll revert today's work and just go back to coding manually. Or perhaps give another chance to GPT 5.2 or Gemini 3? In any case, just like I post when things go great, I thought I should share when things go bad. Or very bad. Basically a lost day & harmfully negative productivity using AI to code.

@ph0rque Haha followed you back now! Thanks! My GitHub is mackdelany

@MackDelany I guess you have to follow me back for me to be able to dm you... once you do that I can send you the link.

For this week's Gauntlet AI assignment, I decided to focus on making a game to teach earthworks planning: transform 3D terrains into building construction sites by precisely balancing cut (removing soil) and fill (adding soil) operations to achieve net-zero off-lot dirt movement.

@_grantsing Irony is everyone that wanted to “stay technical” is now a “manager”

coding is so weird for me now what do you all do during the 1+ minutes that the cursor agent is working? coding used to be locking in for hours typing and thinking constantly with no awkward waiting unless you were building

@yrschrade @Rewkang We evolved to be “humanoid” for a reason. Doesn’t mean it’s best for every task, but it’s probably best form factor for generalising to most tasks. Robot that can do anything & be repurposed as needed is more valuable than your microwave (which is a robot that does one task).

@Rewkang From a hype and funding perspective, I get it. But humanoid robots feel like trying to build a mechanical horse instead of inventing the car. The obsession with two legs, two arms, and a head feels like a legacy idea. Non-anthropomorphic designs will dominate imho

Humanoid robotics is one of the few and obvious 0 to decatrillion opportunities in my life Similar to Bitcoin in 2013, but with larger TAM It's at this point, before mainstream hype that you should have maximal exposure. Few have exposure, and those that do are undersized

@pronounced_kyle @packyM Definitely DO read beyond book 3

@packyM Oh dang, really? I was gonna ask.

Read this trilogy. Read NOTHING about it online first. Just read it. Thank me later.

@VictorTaelin Did it really do that in one shot? 🤯

sonnet-3.5 created a parser for kind2-core in 1 try and it works sonnet-3.5 created a parser for kind2-core in 1 try and it works sonnet-3.5 created a parser for kind2-core in 1 try and it works sonnet-3.5 created a parser for kind2-core in 1 try and it works sonnet-3.5 created a parser for kind2-core in 1 try and it works sonnet-3.5 created a parser for kind2-core in 1 try and it works sonnet-3.5 created a parser for kind2-core in 1 try and it works ??????? it wasn't a "let me start and let it figure out the rest", I just gave it our core haskell file (that has no parser) and it wrote a parser. and it works. and it is beautiful. what is even going on in our planet Earth gist.github.com/VictorTaelin/c…

@_xjdr Also seen that with \n\n. Appreciate that download. Thanks for sharing ! 🤙🏾🙏

vs with control tokens: <|begin_of_text|><|start_header_id|>system<|end_header_id|> You are a happy and helpful AI named Llama3<|eot_id|><|start_header_id|>user<|end_header_id|> Sup Llama3 bro? How you be? Been waiting a long time to chat with you, yo!<|eot_id|><|start_header_id|>assistant<|end_header_id|> Hey there, friend! ears perked up I'm doing fantastic, thanks for asking! I'm always ready to chat and help with anything you need. It's great to finally connect with you too! What's on your mind today? Want to talk about something specific or just vibe out?<|eot_id|>

if I don't prompt Llama3 exactly this way (spaces, newlines, etc.) the responses are a little wild, especially for longer contexts. To me, this is one of the main gaps with OSS models <|begin_of_text|><|start_header_id|>system<|end_header_id|> {{ system_prompt }}<|eot_id|><|start_header_id|>user<|end_header_id|> {{ user_msg_1 }}<|eot_id|><|start_header_id|>assistant<|end_header_id|> {{ model_answer_1 }}<|eot_id|>

@_xjdr @yacineMTB Any prompt / formatting tricks worth sharing? 🤞🏾

I've been pretty meh on Llama3 70B vs opus since it's released but in my time off last week (lawyers are slow) I feel like I have gotten L3 to be able to beat not only our internal models but also Opus and GPT-4o more than 50% of the time on hard multi turn prompts I throw at them. About 800 tricks thrown at it to make it happen but it's an insane milestone. Simultaneously incredibly exciting and incredibly frustrating

@iambrendonjohn Good luck @iambrendonjohn !!

I’m on my flight for LAX->AKL. If we stop in Hawaii, I’ll be returning by sea.

@marchintonnz Oh not sure you have DMs, anyway: flash-24.live/live/stream.ph…

Anyone got a link for a stream to watch Wahs-Broncs in France?

@marchintonnz Will drop one in your DMs

When is the ☕️ club pod dropping @Ifotu42 @eazy_batch12 @finndelany? @YWetzell first guest?

@willstew @Garmin the worst feeling 🤦‍♂️

@iambrendonjohn Took me 3 tries to read it

Last night I finished the first book of the Dune series.

@Leonardo_Hotels I was double charged for a stay this week. What’s the best way to get refunded?

@mullzy2334 💯, never understood why domestic rugby hasn’t trialled the NFL model aka “all on afternoon fiesta”

Afternoon games are also 🔑. Rugby in NZ / Aus needs to take note. Entertainment and convenience are now key factors in modern sporting landscape.

@willstew Once you measure it, the relationship is even bigger than you thought. My RHR dropped 15% in 10 days from elimination diet & no alcohol. And I’m a relatively healthy eater / not a huge drinker.