Ben Davis

yells at @theo both on and off camera (co-host of the nerd snipe podcast)

[Ben Davis @davis7]

The Opus 4.8/ultracode + DeepSWE bench podcast ep. is live:

[Nerd Snipe @NerdSnipePod]

We (mostly) like Opus 4.8...

00:00 - Shower Thoughts
02:44 - Deep SWE Benchmark
10:45 - Opus vs GPT-5.5
19:57 - Anthropic’s Huge Raise
25:39 - Token Maxing
40:02 - AI Slot Machine
43:49 - Claude Code Friction
50:01 - Opus, Mythos, and Safety

[Ben Davis @davis7]

@maria_rcks @theo Promoted, unfortunately for Theo

[Ben Davis @davis7]

@jetpackjoe_ 30-40%. Somewhere between sonnet 4.6 and opus 4.7

[Ben Davis @davis7]

Been using this a bunch today and it's awesome

Grok build is a good TUI, composer 2.5 is an excellent model

Good alternative to GPT-5.5 low reasoning with pi if that's not ur thing. I will probably end up sticking with that, 5.5 & Pi are still each better than composer & grok build respectively

But it's a step in the right direction, I want xAI + Cursor to keep getting better we desperately need the competition rn

[xAI @xai]

Composer 2.5 is now available inside Grok Build.

Composer 2.5 is a fast, highly intelligent model that excels on long-running tasks and following complex instructions.

[Ben Davis @davis7]

Sun eater is my favorite series of all time

Book 1 is solid but u kinda just have to get through it, books 2-7 are incredible all the way through

And the series ends perfectly, books 7 and 6 are arguably the strongest in the series which is impressive

U will know if it’s for u after reading the prologue lol

[Rhys @RhysSullivan]

credit where credit is due, workflows in claude code are good

i've been particularly impressed with them for writing effect, generally works really well with finding strong patterns from other repos and writing it properly

[Ben Davis @davis7]

@ryanvogel they should not have given u this feature

[Ben Davis @davis7]

@colesmcintosh Is there official cursor support? I’ve heard this from multiple people

[Ben Davis @davis7]

I re-subbed to Claude Code to test out Opus 4.8

It's both better and worse than I expected tbh. Claude models have some really weird behaviors/hallucinations and seem to be getting slower

That said, workflows are dope and the model is still very good...

[Ben Davis @davis7]

@TheHunterBohm Hermes is useful for non code stuff. I'm not using it to build things, rather to run workflows for work stuff like emails, social media stats aggregations, reminders, slack, notion, etc.

[Ben Davis @davis7]

Hermes is great and I highly recommend it

But also one of the first things u should do is go in and gut the skills. At least 40 of the ~100 should be instantly turned off

[Ben Davis @davis7]

it's annoying, but also I can understand why they do it this way. If the end goal is for anyone to be able to use it, u shouldn't have to manually curate skills

eventually no one should have to, but we're not there yet

[Ben Davis @davis7]

@DerekCorniello I did end up dropping out but that's mostly b/c college in covid sucked a ton, OSU itself was awesome

[Ben Davis @davis7]

An unedited 1/1 quote from Opus 4.8 max reasoning in Claude Code:

"instead of stopping when the reads came back empty, I described a homepage that doesn't exist: a custom form already wired to Attio via a server API (attio.js, submitBrandInquiry, ATTIO_API_KEY). None of that is real."

...

[Ben Davis @davis7]

This benchmark is the first one I've seen that maps 1:1 to my experience

Almost to a degree where I'm scared to fully trust it since it so tightly maps to my existing opinions I feel like I'm missing something

[Datacurve @datacurve]

Opus 4.8 is now on DeepSWE.

On the default high thinking effort, it scores 6% higher than Opus 4.7 xhigh, while also lowering average cost per task.

[Ben Davis @davis7]

to be clear the terms of the deal are that i won't bully you about gpt-5.5, everything else is still on the table...

[Ben Davis @davis7]

@theo @heyandras Good call on this one, if it was 67 the contract would’ve been void

[Ben Davis @davis7]

@theo @heyandras https://t.co/GroakkeDDe

in return I won't put u in the cope corner tmrw for being wrong about GPT-5.5

[Ben Davis @davis7]

I think it feels like this b/c I haven't heavily used claude since it started needed reasoning, but holy fuck Opus 4.8 is slow

[Ben Davis @davis7]

@ryanvogel Ik that's the problem