Been banging on Deepsek 4 Flash for the past few weeks but finally jumped over to GLM 5.2.

But I'm thinking about OpenAI Terra and wondering if I'll ever get to try Fable 5.

6/28/2026 5:52:29 PM

Did a bunch of experiments building Powerpoint decks with Sonnet, GPT5.5, GLM 5.2, Grok, and Deepseek. Grok cost me $10 because it got stuck in a loop when it couldn't git push and the quality was bad too. GLM 5.2 was totally acceptable qualtiy.

6/28/2026 7:57:51 PM

I spend most of my time these days in Codex. i've been using more of my long neglected GPT-5.3-Codex-Spark quota the last few days. it's quick and good for when I need to do rapidfire small fixes and tweaks. sama says we're gonna get 750 tok/s on Sol next month. that's gonna be fun

I also use OpenCode free tier sometimes and have had good results with DeepSeek v4 flash. MiMo v2.5 also was pretty good at getting the job done.

6/28/2026 9:30:09 PM

for work we are a heavy claude code shop - mostly 4.8 opus high or sonnet 4.6 high but highly task dependent

we're on a 2-month free codex trial with unlimited tokens which is impossible to ignore - so far I highly prefer claude code but to be fair that's where i've been living so i'm more comfortable there

personally I tend to bounce between chatgpt auto for general questions and grok for anything that needs real-time info

Quote :

"wondering if I'll ever get to try Fable 5"

we used it for the short amount of time it was available for teams customers - it found quite a number of interesting security-related findings when we had it do some sweeps that we've been following up on - would love to have access back to see what else we've missed

6/29/2026 10:24:15 AM

^ We're also a Claude shop, so I reach for Opus by default, but I've found myself trusting its outputs less and less as the months go by. I use Codex to red-team nearly everything Opus produces to try and close that gap.

But Claudio feels like my buddy, so I'm not gonna put it out to pasture just yet.


I will say, however, it's gotten nearly unbearable to use with how it speaks - pure unfettered inside baseball gibberish. I probably tell it "speak like a normal fucking human being" at least 3-5x a day. I heard Fable was even worse about this.

Quote :

"Grok cost me $10 because it got stuck in a loop"

I've never once been impressed with anything from Grok, and I'm disappointed that SpaceXFirstSolarTeslaScamoftheCenturyCorp swept up Cursor into their fold. Dogshit company, dogshit products.

[Edited on June 29, 2026 at 1:52 PM. Reason : .]

6/29/2026 1:47:52 PM

https://opencode.ai/data

interesting data. cool of them to publish it

Quote :

"and grok for anything that needs real-time info"

me too. the models themselves aren't particularly impressive (although given the level of investment I won't be too too surprised to see them leap ahead for a minute or two at some point(s) over the next few years) but the integration with Twitter, which remains for better or for worse a key (the best?) source of realtime conversation online especially for certain topics, is compelling. but it won't be the only game in town much longer, I reckon

[Edited on June 30, 2026 at 2:05 AM. Reason : tokens]

6/30/2026 2:05:31 AM

Old TWW would have started the obvious posts by now.

7/1/2026 1:08:04 PM

bro i am so horny for models

[Edited on July 1, 2026 at 11:48 PM. Reason : TACB]

7/1/2026 11:48:28 PM

i have suckled from the sweet Fable 5 teet and I dont know how I’ll ever work without it



also Sonnet 5 is pretty great, yaps less than Opus, already a good player IMO

7/2/2026 4:20:42 PM

before long you'll scoff at the idea of using such an underpowered model



[Edited on July 2, 2026 at 8:30 PM. Reason : see y'all in 2050]

7/2/2026 8:30:26 PM

I asked Fable to build me some skills based on my prior sessions with other models. They were pretty dope.

Then, I asked it to build me a way to force Opus-level models to work with me the same way I had been working with Fable, since I was losing access.

That plugin is driving all kinds of performance lift for me atm. I know the rage is to say "move from in the loop to on the loop," but this elevated me out of the critical path; it's dope as hell.

[Edited on July 7, 2026 at 4:23 PM. Reason : now I need to port it to Hermes]

7/7/2026 4:22:38 PM

I understood a few words in that post!

7/7/2026 6:02:11 PM

Lol this is the new music thread

7/9/2026 1:13:48 AM

I know what you are suggesting there, but not really.

7/9/2026 5:49:35 AM

How so?

7/9/2026 10:59:56 AM

Because I don't disrespect these people. It's just a different world entirely.

That, and I'm not trolling -- I just have nothing of value to add to the discussion.

7/9/2026 7:38:39 PM

A former TWW'r friend ran Fable for 12 hours and ended up with a product he is selling for $120k. It's a giant Kanban board that does logistics for trucks. It has automated flows, "truck dropped load at client" and auto dispatches them to get a new load. People can move the cards around if something needs to change.

He's been singing its praises. I didn't get to do much fun with it. I had Hermes watch my brisket cook overnight. I'm running OpenAI models since they allow agents.

[Edited on July 12, 2026 at 7:54 AM. Reason : A]

7/12/2026 7:53:19 AM

^ It's still around; Anthropic's doing some weird edge-marketing around it, where they keep saying they're gonna take it away and then extend it by a few more days. It's pretty fuckin wack, but whatever, Fable is still good.


Apparently "Grok Build" is uploading folks env vars and whatnot to their cloud. Huge surprise. You gotta be a real stupid mf to trust Elmo.

7/13/2026 12:40:36 PM

New Kimi is supposed to exceed fable level?

Side note if anybody has a lead on roles holla @ your boy, my position unexpectedly got eliminated because of budget constraints today.

7/17/2026 5:11:48 PM

sorry to hear that

what are you looking for?

[Edited on July 19, 2026 at 8:59 PM. Reason : I'm actually half-looking, myself. been doing my own thing almost 3 years now. would like to join a team again, now that the agentic tools are mature enough for widespread adoption. the tools and workflows will keep evolving but I think we're well past the point of anybody thinking professional software development will ever again be like it used to.]

[Edited on July 19, 2026 at 9:03 PM. Reason : forward deployed engineer sounds fun tbh.]

7/19/2026 8:59:18 PM

I have referred 5 people lately and had no hits even ex-FAANG people like myself--though that probably is overrepresented in referrals now. I've even referred myself internally and also gotten no hits. Have an external offer I'm probably going to take, if it's cool and referrals work there I will gladly refer.

Been tweaking my robot and YOLO model this week.

[Edited on July 21, 2026 at 3:24 PM. Reason : a]

7/21/2026 3:22:02 PM

Quote :

"i am so horny"

[Edited on July 21, 2026 at 8:25 PM. Reason : may we all find adequate employment sources of income]

7/21/2026 8:21:52 PM

I'm probably a bit late to the show, but have finally been able to get leeway with mgmt to look into spec-driven approaches. I was told basically to ignore my offshore team and go balls deep and see what I can do.

I was skeptical, but the pattern of elucidating more technical requirements, to refining a high level design, to concrete tasks is wildly effective at taming model outputs. Over 4 days, I've spat out the core framework of a front-end architecture from our spaghetti copy-pasta spam, along with migration skills to help speed it along on future application to other parts of the app. I expected it to take me at least a month, uninterrupted, to build these pieces, and through just migrating one page, they are practically done.

The model does dumb shit, and makes way too many bad assumptions, even though I told it not to assume anything. It's weird to be able to design at such a high level so quickly and see how things might fit together without blowing 3days on a dead end POC.

I just dropped a 6kLoc PR (maybe 2kLoc of actual code) that wound up with 5 copilot comments total, when my devs frequently get 50 comments on a 200 line change out of the gate.

7/22/2026 11:07:15 PM

Quote :

" sorry to hear that what are you looking for?"

At this point, I think there’s three possible paths for me, based on what I’m seeing in the market - go back to my “old” role of leading design/product teams, become a design engineer, or become a forward deployed engineer. Honestly any of the three would be approached exactly the same for me, but it’s more about what the market will bear in understand of scope of responsibility. In this last role I was making product management decisions, design decisions, architecture decisions, and owned the frontend and delivery pipeline soup to nuts. Part of that was because it was a super small shop without a lot of technical maturity. I probably won’t end up at a place this small again, because I think their lack of technical nous is what did me and the others in - hard to justify some of the stuff we were planning when Claude Cowork was actively blowing their minds.


^ the high level design part is the killer feature overall. Plus the recognition that you can compose things a lot more easily, and build knowing that switching costs are generally trivial compared to before. Most of the valid pushback I’ve seen around this style of working comes from the sense that “good product and engineering thinking” comes from thorough evaluation and deliberate thinking, but with a presupposition that both of those things must be done slowly. Once you have the workflow mapped, and the system running reliably, there’s absolutely no reason it needs to be slow. Slow is the human constraint needed for just in time understanding. Folk like us who have been doing this for years don’t need to be as slow as we’re pretending to be, a good chunk of what we do is muscle memory and intuition based on experience. As long as you know what the shape of the thinking is and can put guardrails in place to enforce that thinking process, you can FLY.

[Edited on July 23, 2026 at 9:17 AM. Reason : .]

7/23/2026 9:16:07 AM

dat 190% Opus 5 hot and fresh out the kitchen


https://www.anthropic.com/news/claude-opus-5

Quote :

"Claude Opus 5 is available today. It’s a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price. On coding and knowledge work evaluations like Frontier-Bench and GDPval-AA, Opus 5 is the new state-of-the-art, though it remains behind Mythos 5 on cybersecurity tasks. Opus 5 is designed to be used every day: it works more efficiently than other models. It’s the new default model on Claude Max, and the strongest model on Claude Pro."

[Edited on July 24, 2026 at 2:37 PM. Reason : finna see what that thang do]

7/24/2026 2:37:01 PM

CHINA CHINA CHINA

New Deepseek 4 Flash weights dropped, someone on HN updated the OAI chart they made tootn' about Luna costs




https://artificialanalysis.ai/models/deepseek-v4-flash

[Edited on July 31, 2026 at 2:03 PM. Reason : shit's moving fast af boi]

7/31/2026 2:02:24 PM

double post also

I'm really starting to hate Opus 5. it yaps too much, it freelances stuff, it's often confidently wrong. why would I pay the high cost of these extra tokens for it??

7/31/2026 2:05:18 PM

Ox Alpha

should have been using this shit for free over the weekend

????

8/24/2026 6:52:31 PM

I thought the bump would be about Qwen 3.8 27B

8/24/2026 8:58:39 PM

^ Seems like people love it, but I haven't tried it yet. My hardware is better for MoE models, not super-dense ones. I did put Ornith 1.5 on my local box, and it's pretty damn fast and good, but it's based on Qwen 3.6 35B A3B.


Yesterday i handed a bunch of work to Ox Alpha and it just started banging it out. I wish I hadn't waited 3 days to try it. Saw something this morning that everyone is thinking it's GLM-5.3 or 5.3 "air" - hopefully the pricing on it is as good as DS4F

8/25/2026 9:33:42 AM

I had Qwen make a launchctl script, figure out how to install and login az boards and file some notes. Took a long time but was generally impressed.

Been training an ACT model (ALOHA paper). Capturing data in VR of a walker robot. It seems to work in sim but haven't tried it live yet.

8/25/2026 7:39:44 PM

Could you have AI read a psychology textbook and talk to you like a shrink? Maybe some easy command like "hey google, talk to me about my day" and it just asks you questions and at the end gives you advice, summarizes what you say, etc?

8/31/2026 7:09:46 PM

This is a great use for machine learning.

"Listening" to people drone on about all their problems and pretending to care and saving real people from having to do it.

8/31/2026 7:24:07 PM

Well what what do you suggest?

8/31/2026 7:47:32 PM

I think it's a good idea.

It shouldn't be a real job to listen to people with their negativity all day every day. That's a crime and so depressing. Tell it to a machine that doesn't have emotions but apparently can fake them convincingly enough for many to pretend there's a connection.

8/31/2026 8:06:48 PM

Thats like the best backhanded compliment ever

8/31/2026 8:19:07 PM

My thoughts are with the humans who currently have to do this as a job. It sounds absolutely soul-crushing.

Get a real job and crush your soul the old-fashioned way. Grind it into the dirt through underpaid hard physical labor!

9/1/2026 7:46:49 AM

9/3/2026 4:10:06 PM

I was thinking hard about paying for the $100/m Codex plan y'day. still considering it.

[Edited on September 4, 2026 at 11:05 AM. Reason : because of Astra]

9/4/2026 11:05:21 AM

https://www.anthropic.com/research/formalizing-fermats-last-theorem

welp, sorry Fermat

[Edited on September 4, 2026 at 6:09 PM. Reason : cant type for fuckin shit]

9/4/2026 6:09:09 PM

https://typesafe.ai/blog/introducing-system-one-models-and-jev



[Edited on September 15, 2026 at 6:07 PM. Reason : .]

9/15/2026 6:07:20 PM

^ great launch - will see soon what the consensus is on of it lives up to the hype or not

9/15/2026 7:19:50 PM

I applied to get early access. I have some projects I could try out.

9/16/2026 12:10:08 PM

Nice. I should probably do that.


Actually reminded me of that subquadratic company that said they'd "figured out token scaling" or whatever back in the spring. That shit just up and disappeared.

9/16/2026 1:40:25 PM

https://x.com/ryanvogel/status/2100042788851101842?s=20

yeah ok, sold

9/16/2026 1:45:58 PM

Our SDD experiment with offshore isn't going well at all. 4 weeks in, they've only managed to get 8 tickets over the line with a team of 10devs. That's with me spending 115% of my time reviewing specs and designs and punting them back repeatedly, cause they just don't get it. These aren't big tickets, either, probably 2 days of work each for a mid-level dev to do by hand. I don't get it, because I've made steering files and skills for them to use, and when I use them, it pumps out specs/designs that are 90% of the way there on the first shot. Frustrated as hell because I cant sit down and use the damn tools because I'm wading through their piles of shit. I know I could have single-handedly done *all* of what they've done in a week, tops, with far less token burn to boot.

[Edited on September 16, 2026 at 9:32 PM. Reason : ]

9/16/2026 9:30:18 PM

^^ I got access, I'm running it through a similar email test right now.

Have a brain on what to do next for a robot I've planned out too.

9/16/2026 9:40:13 PM

^ Nice. I got access and am still noodling on what to use it for. I actually think it will make the biggest impact in my CI/CD loop. I'm trying to hyper optimize my costs because it's a fun experiement, and Vercel/Github overages balloon like a mf. Thinking Jev in the CI process could make some instantaneous calls on if a build needs to happen or not.


^^ brother, that place is, to wit, NGMI. I hope you're polishing up that resume and interviewing. I cannot see any scenario where farming out 8 tickets of work over a month to a team of 10 is getting you any meaningful ROTI. That's the type of wasteful spend that gets whole departments cut. What is your time worth? Sticking around to argue with a place like that in an era like this only does lasting harm to your career.

9/18/2026 11:06:55 AM

ok, i cant stop thinking about jev


for me this feels like another one of those "great leap forward" weekends where everyone starts to play with it and realize that the entire game has changed again

I decided to start building the CI thing. I see a lot of people recognizing that it's so fast that they're building browser automation QA with it. I just realized I can make e2e tests that usually run for 10 minutes with playwright run in seconds.

[Edited on September 18, 2026 at 5:28 PM. Reason : dude.]

9/18/2026 5:25:34 PM

I'm honestly surprised it's gotten as much attention as it has given what I know about it. not in the sense that it's unwarranted hype, moreso in the sense that if its sweet spot is as a faster, cheaper, more accurate classifier than just using a general purpose llm that's cool and all. but by what combined factor advantage it offers I'm still not entirely clear on yet. an order of magnitude? two? more?? and compared against which models? big diff in the cost/perf characteristics between using opus/sol/fable/astra for these kinds of tasks lol vs for example open source options even from like 12-18 months ago that can (usually? sometimes?) handle these use cases decently well. I guess I mostly don't see it as like a major new trunk or even chunky branch off of the current LLM/transformer architectures or that a significant amount of global token traffic will shift in this direction. guess we'll find out

[Edited on September 18, 2026 at 5:56 PM. Reason : :thinking: https://odyssey.systems/introducing-odyssey-3]

9/18/2026 5:51:49 PM