Jeremy Chone

#ProductionCoding - Rust - Cloud App - AI Production Coding - - -

[Jeremy Chone @jeremychone]

Think Big App, Small Model: GPT Luna vs. Terra vs. Sol

Defaulting to biggest model can 5x to 10x AI Cost.

👉 Think Big App, Small Model
👉 Model escalation is the exception
👉 Continuously flatten the complexity curve of your system

#ProductionCoding

https://t.co/jSLa5LdTtb

[Jeremy Chone @jeremychone]

@AjaySinghadiya8 I've used v4-pro for production coding (#rustlang) for months.

It's good, but Luna is close, if not better.

Those v4-... models reason a lot, and reasoning costs $$

Luna's speed is a game changer

I'll try the new v4-pro, but not for "open weight", since it's no longer cheaper

[Jeremy Chone @jeremychone]

@ns123abc Luna cheaper

[Jeremy Chone @jeremychone]

Significant price increase expected

DeepSeek's cheap days may be over

DeepSeek v4-flash was already pricier than Luna on real tasks

It'll be interesting to see v4-pro's price

I am aaig DeepSeek fan and user, but OpenAI Luna my first choice for now (mostly #rustlang coding)

[Jeremy Chone @jeremychone]

@Da7_Tech but 4x to 8x slower than Luna.

Luna got webgpu/sharder of my benchmark first go.
v4-flash, did not.

but still good to have those open models putting pressure.

[Jeremy Chone @jeremychone]

@westxsh @theo Yes, way overdue.

Next one will be about 3 Levels channel pattern

Level 1 - app wrapper api (on top of crossfire)
Level 2 - type alias on top of level 1 (e.g. TuiTx, TuiRx)
Level 3 - type wrapper on top level 1 (e.g JobQueue, ...)

Hoepfully this weekend ...

[Jeremy Chone @jeremychone]

#AIPack v0.8.33 released

+ Models
- new #QwenCloud provider w/ qwen3.8-max
- Luna, Terra, Kimi k3 cost update
+ aip install packs from git repo
+ TUI: Loop group display
^ API: aip.cmd.exec with .cwd

[Jeremy Chone @jeremychone]

@Da7_Tech But too slow for me. Take 5 to 15 minutes per run (cloud app, Rust coding)

How do you have it faster?

[Jeremy Chone @jeremychone]

@theo I think they are "vibe coding it."

[Jeremy Chone @jeremychone]

@mattpocockuk To be honest, I never got this grill-me thingy.

I do

chat md -> goal md -> plan md -> managed loop

Then put/update spec mds with goal/plan when needed

Goal, plan, loop are optin

[Jeremy Chone @jeremychone]

@Da7_Tech Honestly, I use APi for all, at the end, cheaper

[Jeremy Chone @jeremychone]

@Da7_Tech Luna is cheaper or same price and so much faster.

Been a big fan/user of dflash for months. But Luna new pricing is game changer.

the Luna speed for similar dflash price is crazy

[Jeremy Chone @jeremychone]

prompt:

[Jeremy Chone @jeremychone]

Qwen 3.8-max is good,
but so SLOW, and NOT cheap
(from qwencloud .com)

For this demo, WebGPU, 2 prompts (see in the comments)

output: 63k tokens (including 56.5k tokens)
price: $0.40
time: 20 min !!!

Luna-max was $0.06, 1.5 min, for a similar result

👇 prompt in comment 👇

[Jeremy Chone @jeremychone]

@fabrice_mayrand go Luna, cut 70%

[Jeremy Chone @jeremychone]

Playing with WebGPU / Shaders for my next AI benchmarks

Kind of cool, but not really what I asked for.

Tuning...

Had to use Sol to finish Kimi K3 work.

K3 did good though... but so slow ...

[Jeremy Chone @jeremychone]

@Da7_Tech Agreed. With the right flow, the price structure is a game changer.

I've been doing some pretty big stuff with Rust, and it's surprisingly good.

And it's so nice to have a 10-run loop for $0.30.

[Jeremy Chone @jeremychone]

@VadimStrizheus Unknown opinion:

AIPack/pro@coder > Codex > Clause Code

[Jeremy Chone @jeremychone]

@OfficialLoganK In fact, we bet on the harness.

It's all about compounding.

Models don't, harnesses do.