Donald Ness

Was doing a live product demo for a customer. Had the dev server running in a Claude shell w/ context I was demoing. A new feature crashed the server (config error) halfway thru, Claude detected the problem, updated the config, and within seconds the demo was back on track!

@NielsRogge Reflection 2.0?

After checking his LinkedIn, the chances of it being a scam went up subquadratically

Introducing SubQ - a major breakthrough in LLM intelligence. It is the first model built on a fully sub-quadratic sparse-attention architecture (SSA), And the first frontier model with a 12 million token context window which is: - 52x faster than FlashAttention at 1MM tokens - Less than 5% the cost of Opus Transformer-based LLMs waste compute by processing every possible relationship between words (standard attention). Only a small fraction actually matter. @subquadratic finds and focuses only on the ones that do. That's nearly 1,000x less compute and a new way for LLMs to scale.

@_agarner Primarily terminal. Much of the time from my phone via getmoshi.app. But when at my desk I’m reviewing closely in the IDE and guiding it, updating docs and instructions to push it in the direction I want. I have strong opinions on coding, design, and architecture.

@programmarchy Do you work primarily out of terminal or in IDE? I’ve been all terminal but see some using ide

Paying for Claude Max has made me more productive because if I’m not maxing out every five hour window I feel like I’m leaving money (tokens?) on the table. Overnight is still underutilized because I haven’t gotten good enough at planning. I’m also still very hands on directing.

@karrisaarinen Yes I feel like this constantly

A common dynamic I observe with AI: it feels most impressive when you don’t know much about the subject, don’t care or don’t have a clear idea of what the you want. This applies across design, code, legal, and more. If I don’t know code very well, every piece of code it writes feels very impressive. Once you know what something should feel or look like, it becomes almost impossible to guide AI there. And you definitely can’t one-shot it.

I have changed my mind on how AI will impact jobs in America. Previously, I believed AI would replace many entry level roles typically filled by young employees. The technology would then work its way up the organization and eventually reduce the total number of jobs in a company. The data is saying something different, so when I get new information I am willing to change my mind. The number of software engineers being hired has been increasing. The number of open software engineer roles is growing. The number of new college grads who get hired has increased 5.6% over the last 12 months. The unemployment level for people aged 20-24 years old who have a college degree has fallen from nearly 9% to almost 5% as well. The Wall Street Journal recently wrote “AI created 640,000 jobs between 2023 and 2025 in the U.S., according to an analysis by LinkedIn of job posting data, including new white-collar positions such as Head of AI and AI engineer.” And I am starting to see companies throughout our portfolio aggressively hiring to keep up with the demand for their products and services. If AI can make employees more productive, which is widely accepted as fact, then companies are going to want as many productive units of labor as possible. This is a key reason why I am changing my mind. AI appears to be a magical technology that will make companies more productive and more profitable. The net result will be more corporations, more startups, and more jobs. All three are big, positive wins for the American economy.

@zeeg @grahamneray How does authz fully solve the problem though? In a multitenant environment you still ultimately need to filter the data returned by the db. RLS policies simply require that filter to be in place at the data level. What am I missing?

auth is not that hard, and its certainly a lot easier when you can write software and test software to support it if someone cant write some permission statements and learn baseline of auth and permissions they should not be shipping software this is a fundamental requirement of building software at most companies you dont have to be an expert, but you have to be competent part of the problem is in the last decade we've outsourced just about every problem under the sun and no one learns the fundamentals. it doesnt mean that there shouldn't be supporting software for a lot of problems, but you cant use a calculator if you dont learn math

RLS was a mistake and folks exposing that level of complexity to less technical users is asking for trouble. It was a mistake in Firebase. It’s a mistake in Supabase. It will be a mistake in the next product too. I personally - even knowing how to secure it - would never touch it. It’s the worst security footgun you can imagine. One small mistake and your data is available to the world.

@james_reyes One or two agents on separate repos so far. Just read about agent teams and /batch now after hearing it from you 🤯 Big recalibration needed from my experience ~6mo ago when parallelizing didn't seem plausible.

@programmarchy Ah got it, do you find yourself parallelizing much? I use Ghostty and have a few agents going at once. I've been playing around with Claude's agent teams and using /batch to parallelize tasks when not much planning is needed.

Over the past year I've switched between Cursor and Claude Code twice. I've recently switched back to Claude primarily because of Remote Sessions. Being able to continue coding sessions on my phone while I'm away from my laptop is gold.

@james_reyes I’ve used parallel agents where you spin off a few models on the same task and then compare the results to pick which was better. Was interesting but felt like lots more work for marginal improvements. Maybe there’s a smarter way to use that feature. What tools are you using?

@programmarchy Ah got it. Since you're using them again, and out of curiosity, do you use Cursor's multi agent features? Haven't heard as much about them since they launched it, wondering how it stands up in practice

@tannerlinsley plus we’ll all get a tan……….. thank you i’m here all night

Instead of doing a TanStack Conf, I'm seriously considering inviting everyone to just meet up on a Caribbean-bound cruise ship. It'd be cheaper in every way, you could bring as many family/friends as you want, and instead of wasting time on talks/booths, we can just chat endlessly about JS/TS/Web over unlimited freestyle soda machines, soft-serve ice-cream and mini golf. TanStack Cruise 2028

@Layton_Gott If I were new, I’d follow the lazy smart path and start with AI. Instead learn engineering principles like carefully designing specs, thinking through edge cases, organizing systems, and figuring out ways to make code testable.

Learn to code first, then use AI? Or Start with AI, learn as you go? Which is better?

@melvynx @martindonadieu damn I fell for it too

@martindonadieu ❤️ i lie so you can react

Why you should NOT use Supabase: 1. SDK Trap - not TypeScript by default, type-gen is bad, easy to make mistakes 2. Security with the SDK is terrible - need to set up a lot of things to have something work 3. Vendor lock-in with the Auth, don't have your keys 4. Pricing way more expensive than a $8/month db + better-auth 5. Better alternative for all tools they provide

@melvynx Dunno. Supabase is nice. Great DX especially for local dev, nice web UI for DB access, too. Supabase Auth is new and you aren’t forced to use it so I don’t see how it’s vendor lock in.

@boringmarketer Updates keep breaking stuff for me and it’s been a lot of work to keep fixing it so yes. It’s a pivotal project for agents but the bigger labs are quickly surpassing it in terms of product polish and ease of use.

anyone feel themselves using openclaw a bit less lately? personally I feel like it’s memory is not great and it’s tool use even with skills is not close to claude code I like it for quick on the go things but it’s not a workhorse for me by any means at this point

Had some Claude Code remote control sessions hang. Found out about Moshi thanks to @odd_joel. Beautiful app that has remote control at the terminal level, packed with features like voice dictation, and generous free tier. Love finding apps like this.

@odd_joel Wow the app looks so good. With voice dictation 🤯 This is awesome man!

@programmarchy yeah losing your session mid-work is the worst, hope moshi helps with that

@odd_joel That’s a good idea! I have had my CC remote control hang a couple times which was pretty frustrating since I didn’t have a way to access the terminal from my phone.

@programmarchy remote sessions are great. for direct terminal access too, mosh protocol is a nice complement — your session survives phone sleep and network switches so you never have to reconnect. been using Moshi for this, also has push notifications when tasks finish

@james_reyes Mainly because Claude went way off the rails a few times and I went back to Cursor so I could get access to other models.

@programmarchy What brought you back to Cursor previously? I hear a lot of people moving Cursor > CC or Codex (or newer orchestrators) but not in reverse

@advaitpaliwal This is such a nice project to learn from. The code is very well organized and easy to read. I recognize pi-ai from OpenClaw!

I built Feynman, Claude Code for research. I gave it a question and it came back 30 minutes later with a cited meta analysis. It can also replicate experiments on Runpod, audit claims against code, and simulate peer review. Open source & MIT license, link below