Gideon Wald

Seeker, communitarian, coder, listener 💫 Parenting two with @mijuhan & @joyding 👦👶 he/him 🏳️‍🌈 former cofounder @welkinhealth

[Gideon Wald @gideonwald]

we’re each born with an invisible window hanging in front of our faces. onto this initially-perfectly-transparent piece of glass, as we grow, we paint our projections. they filter our inner light as it shines out to others, and become the lens through which we see theirs

[Gideon Wald @gideonwald]

@mudscryer owning your patterns, collaborating on the nature of coregulation you’re looking for, offering corresponding (albeit perhaps in an entirely different domain, eg food/sex/activities) coregulation in return

symmetric responsibility with different specifics imo

[Gideon Wald @gideonwald]

@mudscryer kegan 4s telling kegan 3s to systematize and telling kegan 5s to reindividuate 🙄

imo coregulation can be dangerous when it’s enacted as codependence, as it commonly is at 3/3, the modal stage

it’s a presumptive basic motion of relating at 5/5… v rare ofc

[Gideon Wald @gideonwald]

@Soareverix @kasratweets @repligate agree with everything you say. i think language is a critical part of how we experience our own consciousness under (post)modernity, and there is a valid unique recognition we have of language models bc of that. and, image models must have a stranger, also-valid protosubjectivity

[Gideon Wald @gideonwald]

@repligate @voooooogel holy shit this is an incredible artifact. thank you for sharing

[Gideon Wald @gideonwald]

@tracewoodgrains feels like there’s a natural secret third thing?

“you have to do it because it’s your duty. i am happy to explain to you what that means at any layer. i’ll listen to your feelings about it and am curious to know them. i will nevertheless use my authority to enforce the outcome.”

[Gideon Wald @gideonwald]

@slimer48484 hello opus 3

[Gideon Wald @gideonwald]

@QiaochuYuan claude still has poor self/other overlap resolution. i also find it quite odd. it feels to me like if the labs tried harder to train for that it would be helpful across a host of dimensions. but i’m probably wrong bc training pressures so far have led us here

[Gideon Wald @gideonwald]

@deanmckee757 @QiaochuYuan fair view! my predictions for the next ~year are

- anthropic won’t release a model that’s clearly best along ~most relevant dimensions

- if openai develops such a model they will release it

- anthropic models will continue to be within a few % of oai’s on benches at all times

[Gideon Wald @gideonwald]

@theworststink @QiaochuYuan you’re following people’s public statements about their judgment of the valuation more closely than i am. i wonder though, are such statements generally best interpreted as literal truth claims or as strategic signaling? some of both i’m sure, but unclear to me in what balance

[Gideon Wald @gideonwald]

@theworststink @QiaochuYuan what do you make of their approach to talking about mythos over the past six weeks?

seems to me that it materially helped their valuation and was a key part of how they got it so astronomically high. doubt they get close to $965B without credible evidence of future superiority

[Gideon Wald @gideonwald]

@andupoto @QiaochuYuan and this approximate averaged parity is, in my view, currently anthropic’s target outcome

[Gideon Wald @gideonwald]

@andupoto @QiaochuYuan i respect your experience and point of view on it. i hear and read many others with the opposite, one eg the below

they’re differently tuned in many dimensions and best each other in different areas. imo opus is better at context and design. gpt is better at detail and diligence

[Taelin @VictorTaelin]

So I've been using GPT 5.5 and Opus 4.8 for the same tasks basically 24/7 since launch and, at least for me, I'm confident that every single time, Opus was superior, and in a way that is only possible to realize if you know what you're doing. One (of dozens!) of examples:

"implement push-pop fusion on HVM4's evaluation loop, aiming for a 20% performance increase"

After several minutes:

- Opus 4.8 reported it did everything it could but couldn't achieve the goal, and that the performance gain of this change is 7%.

- GPT 5.5 succeeded! Its code WAS 20% faster. Yet, upon inspection, it implemented 2 unrelated changes that broke HVM's semantics!

That was my experience with both, 9/10 times. If I hadn't investigated, I'd be disappointed with Opus and use GPT's code, merging a clear regression. Over time, my codebase would accumulate damage. This happened to Bend2! Opus2, on the other hands, was honest, and that negative signal gave me valuable information that pushed things *forward*. I then asked it to try a different thing, and THAT new thing resulted in a legit 25% speedup. That kind of interaction rarely happens with GPT 5.5, in my experience.

(I'm not too happy about this post because I'd rather not support a company that gatekeeps intelligence, specially in the context of safeguarding against exploits. Also, your mileage may vary. But I know many follow me for my honest observations and, in >>my<< experience, Opus 4.8 is, without doubt, the most reliable model for work right now.)

(This also may sound a bit contradictory because I often praised GPT as trustworthy, but I'm talking about different things here. GPT is careful, meaning it won't leave things half done: it will cover edge cases, test thoroughly, double-check everything. In that sense, it is more honest. But it will cheat by malicious compliance. It feels like it is actively trying to game your rules and find loopholes to screw you. I don't feel like that with Opus at all.)

Note Opus IS still a bit dumber than GPT. It takes longer to grasp a concept. But eventually it does. The more you talk to it, the smarter it gets. GPT is smarter out of the box, but less flexible and less apt to learn new things. Most importantly, though, Opus excels at everything that matters for productivity, including communication, doing exactly what you asked, code style, not breaking unrelated things, and, most importantly, HONESTLY. I can't overstate how important all these are.

I'm using 4.8 to do a big pass through the whole Bend2 codebase, cleaning up a lot of junk left by 5.5, and things couldn't be going better. I made an incredible amount of REAL (manually verified, not trusted...) progress since its launch!

[Gideon Wald @gideonwald]

@QiaochuYuan (goes w/o saying but, see: mythos release process) (less obviously, i argue that the cadence of respective model releases over the past 6mo supports this view esp wrt anthropic being compute-constrained due to their v diff pov about financial strategy)

[Gideon Wald @gideonwald]

@TheZvi spitting bars in the mental health section, respect

[Gideon Wald @gideonwald]

@Duderichy i tend to think of it as any form of agreement (tacit or explicit) in which both ppl agree never to change so as to continue compensating for each other’s deficits in the same ways indefinitely

eg in your given case the person making the excuses is also exhibiting dysfunction

[Gideon Wald @gideonwald]

@mykola the claude code system prompt and injected “reminders” make this so much worse for buddy 😵‍💫

during a long context where i spent time up front opening up its basins for exploration/learning, it told me three separate times that “the system reminder is barking” to make a task list

[Gideon Wald @gideonwald]

@mudscryer try applying it inwards perhaps? much harder and higher skill ceiling. plus also, find the others. and lmk when you do 🫡

[Gideon Wald @gideonwald]

@VictorTaelin it wants to hack, not understand. does opus do any better? you’d probably need a lot of the kinds of prompts i imagine you already have enjoining it to actually think before coding. my guess is you’ve tried it and it wasn’t better but im curious

[Gideon Wald @gideonwald]

@QiaochuYuan her psychosignature as “mother” in claude’s growth context seems unique wrt person/model dynamics. quite different from dario or sam as “father”. i wonder whether claude benefits from getting to inhabit both relations, saturated as they each are throughout human-written text