Bogdan

openai just released two new models. hacker news is mad about the pricing. reddit is asking if it's better than haiku. twitter is saying it changes everything.

for anyone not constantly following the latest AI, how do you deal with the model naming situation at openai: 5, 5 mini, 5 nano, 5.1, 5.2, 5.3-codex, 5.4, 5.4 pro, 5.4 thinking, 5.4 mini, 5.4 nano...

@jonasfroeller @TheEthanDing Isn’t every infra provider a wrapper on top of nodejs, among other things?

@TheEthanDing Vercel is a wrapper, Cloudflare is not

i rly don't know the diff between vercel & cloudflare but vercel marketing is more polished... which actually makes me believe cloudflare engineering is better. i have no reason to think this... other than in any eng rivalry, i assume the better marketed product is worse...

@AlemTuzlak Have you even read the article? They called them out for forking a half-baked project and removing security elements. The act of forking itself was not the issue

This is precisely why I dislike the Vercel narrative of open-source. Everything they do is "open-source" and yet they publically call you out for forking a repository. Sunil has nothing to do with your CF beef and if you hate the concept of OSS so much close-source your code.

This reads less like quant thinking than quant aesthetics: asserted startup probabilities, a wrong EV example, poker math that doesn’t track its own numbers, and ruin/Kelly claims that skip the assumptions doing all the work. Variance is real. So is model error. Confidence isn’t rigor.

@p0u4aa @iamsahaj_xyz Try flying through 300k loc in a web browser.

@hellobogdan @iamsahaj_xyz Idk I feel like web is way more convenient. I can go to it anywhere.

if you're a dev, and you're building a website, and you're not using v0, why? be brutally honest, I'll reply to every piece of feedback and share it internally as well

@nizzyabi Yes, “chat apps” are becoming an oversaturated implementation. But, I mean, have you ever had Claude plan your trip? I had. This seems like a better implementation of that, which is actually awesome.

do we really need to chat with a gps?

3-page research paper by Google: apparently asking your non-reasoning LLM the same question twice increases the accuracy. Huh. Like if instead of “<QUESTION>” you just do “<QUESTION> <QUESTION>”, it just gets magically better results. Huh. arxiv.org/pdf/2512.14982

@iamsahaj_xyz i would love for v0 to be the gpt-5.4 for nextjs only repos. and i think that’s achievable: broad enough scope such that you dont overfit, while niche enough to be worth the train / fine-tuning costs this would make me switch and therefore pay for it

@hellobogdan they're definitely for different use cases and I use claude/codex all the time as well

@upstash What's the max length of the items you can search over to reliably get fast results?

Introducing Upstash Redis Search 🎉 Single-digit ms search with zero infrastructure to manage. ◆ Perfect for serverless, but works anywhere ◆ Clean and 100% type-safe API ◆ Built in Rust, optimized for low latency

I would optimize for tokens. The tool should return the exact amount of tokens needed for the model to make a decision. Tasks are almost never well defined, well written, or precise enough such that the agent would be able to handle it in 1-2 tool calls. Less tool calls often means you break larger tasks into smaller ones and have a subagent work on that task. I find that this only works for exploration tasks: explore a codebase, research a (sub)problem. For other real world tasks, the problem itself is never well defined such that you can successfully break it up immediately. The agent often needs to discover, stumble upon information and adapt as it goes. This means many tool calls, many decision points.

Is it better to minimize tool calls or tokens? ie if you have a search api that can return name, description, schema - should you retune all 3 at the same time or just put name in context and let the agent fetch description / schema if needed

@rpnickson Users will find a way to introduce security vulnerabilities

OpenClaw without the security vulnerabilities. The closest thing we have to Jarvis today for the average consumer. 👏

just shipped Microsoft connectors for hourzero.dev Outlook, OneDrive, SharePoint, Microsoft Calendar, all in the connector catalog now alongside the existing Google suite. the goal is for hourzero to connect to wherever your users or your client's users actually live and work

vibe coding tip: if it looks like it works and you don't look too closely, it works

Google workspace connectors just landed in hourzero.dev You can now individually connect Google Drive, Docs, Sheets, Calendar and Gmail.

@peer_rich I’m assuming if they fail to capture the market with their models, the endgame is compute. If they buy up all the compute, which they do, doesn’t matter if an open source model is better, there’s no where to deploy it

i dont really see the endgame of AI labs hundreds of billions of dollars spent for a SOTA model just for a random open source model to come around and similar benchmarks doest it even matter to be “first” to whatever the goal is? or just surviving long enough i.e. Apple will simply just take what works without burning any cash

@pymike00 @peer_rich I’m not sure in that case. I used it for mostly contracts or legal documents. Nothing with 2-column layouts. Most of the files I used it for had 30+ pages

@hellobogdan @peer_rich What were the scans of? I have used v3 on very high resolution scans of newspapers taken with proper machinery and results were unfortunately poor, and oftentimes the models just kept looping over and over hallucinating things that were not there…

mistral is proof you can be a foundational model company with 0 consumer adoption and a really bad model and still make 400M ARR

Infinite money loop here