Dominik F. Farr

@daniel_mac8 @synthwavedd Well, they can’t focus on Gemini CLI users because they don’t have any

@synthwavedd The fact that Google cares about the benchmarks so much is what's holding them back.

🚨"We have a new version of Gemini coming very, very soon", confident on benchmarks - CEO of Google Cloud

@nickbaumann_ It did mine weeks ago. It did an absolutely great job researching all regulations and reasoning about deductions, the only hard part was to actually fill out the pdfs. Took a lot of iterations to have it place everything on the right lines. How did you do that last part?

Codex did my taxes today. Thank you Codex.

@polynoamial A shift in thinking that many people miss. No one really wants to build AGI. We want to build the thing that builds AGI.

What we really need is a benchmark where AI models make AI models that play poker.

PDFs suck. Tax returns should be JSON to an IRS POST endpoint.

@gabriel1 When the machines take over you will find yourself in some black mirror episode rewriting the same code for infinity. I always make sure to say thank you to codex.

i gaslight codex for a living. it makes me some beautiful code, but i tell it that it's the most disgusting complex code i have ever witnessed. and somehow it improves the code further, making the most consumable code in existance

@dylan522p Everybody knows it sucks. The only reason to still use it is if one is some ancient company that uses Active Directory and all the other MS stuff. You are running a startup, so just walk away from them.

Microsoft Office 365 is such a dog shit product. Literally been breaking for the last 2 weeks when we try to buy more licenses. We use Gmail and Slack, so it's silly to even use this. Sharepoint is buggy a f too. Why no one else make a functional collaborative file systems :(

@deredleritt3r @signulll I would not be surprised if a part of Demis hopes that LLMs don't work out. It's a problem with a lot of the old Gard of researchers: LLMs are not supposed to work because they aren't ingenious enough in their design. Classical Yann LeCun syndrome.

I find the below video instructive. Dario Amodei opens with explaining how Anthropic plans to reach RSI through automating AI research, which is reached through Claude Code. Demis Hassabis basically says that, if this doesn't work out, then the path to AGI is through world models, robotics and continual learning. Implying that he's focused on the latter direction, and not particularly interested in the former. The fact that OpenAI and Anthropic are releasing monthly improvements to their coding harnesses and models, but Google is completely absent from this race, confirms my suspicions empirically.

if true this would be absolutely insane. (there is no way this is true). can someone please explain?

@NaderLikeLadder @DanielLurie Don't ask for permission, tell the AI to solve the captcha. That's a much more scalable solution than writing cli's for every old government website.

Hey @DanielLurie 👋 my openclaw can’t pay my parking tickets because of the CAPTCHA. If I built an SFMTA CLI, would you let us use it?

@firstadopter I try Gemini cli every few weeks. Last time it still had actual bugs, as in it would randomly crash. @OfficialLoganK can we please just lock @JeffDean and Sanjay in a basement for the weekend and ask them to do a rewrite?

Does Gemini have anything competitive with Claude Code and Codex? I don’t see anyone talking about it.

@karpathy @rasbt @teortaxesTex Old thinking. You just have to enable your agentic loop to look at images.

@rasbt @teortaxesTex Nice! My autoresearch would love some markdown version of this - pool of ideas.

I (finally) put together a new LLM Architecture Gallery that collects the architecture figures all in one place! sebastianraschka.com/llm-architectu…

@lafaiel First step would not be Wildcat Lake but Arm. x86 is a flawed legacy architecture and deserves to die. To much baggage, too many design elements just due to backwards compatibility. Intel should finally move on to RISC.

Good thing about Apple is they put pressure Laptop makers are now under pressure to make laptops with Wildcat Lake, 100% sRGB panels, and aluminum enclosures for $599. If they can't handle that, they'll start losing entry market share to Apple

@cullen_mi @pashmerepat All raw

@Dominik2718 @pashmerepat Do you give it all your docs raw, or redact PII?

Everyone talks about Codex for coding. I want to hear about the other stuff. If you're using it beyond writing code, what's your workflow?

Put all documents (W2s, broker docs, contracts for self employment income, last years taxes, …) into a single folder/repo and ask Codex to download all IRS forms (and state tax forms if state income tax exists) and then just fill it all out. Advice it to use playwright mcp to open the pdf so it can do visual inspections using screenshots. It’s gonna still be a bit of an interactive process, so treat Codex as you would any tax advisor.

@Dominik2718 @pashmerepat I am curious about your taxes workflow. Personal or business? I have been thinking about this as well lately as I am trying to automate all my bookkeeping and get rid of my CPA

@steipete @openclaw How do you access Discord from codex? Is there an MCP/Cli?

Since PRs at @openclaw are basically reverse entropy, I'm now using codex to run data analysis on Discord to filter out the most important pain points to see where to work on next.

IMO people still think of codex as a tool for coding, when really you can do all kind of data analysis/work there.

@johnennis If just Gemini CLI wouldn’t be as bad

Pour one out for Gemini 3 I think the Gemini models are really underappreciated for their creativity, especially on the visual side The greatness of Gemini 3 is what makes nano banana pro so awesome

@jsuarez Give JAX another try. It has a backend in the sense that you can read and debug the intermediate language it compiles to. Also helps if you’re working with people that like inheritance and abstractions, cause this shit is just much harder in JAX and so they stop trying.

We have inherited a colossal mess from Torch. Their low-level backend is an unmitigated disaster (still better than Jax etc though because at least they have one). The new backend would be around 3500 lines with torch but only 4000 in raw cuda. So torch barely saves you anything and introduces a dozen new places for errors. Plus at least triples compile time. The only downside that does balloon up the code a bit... there are a few ops where you really, really want Nvidia's cudnn kernels. The interfaces to these are bloated because they are wrapped as opaque structs. So you get +100 lines of boilerplate to call a conv. But torch etc. are not "simplifying" that at all. They are just adding 5 layers of wrappers instead of 1. I actually don't know if you really even want convs for non-image grid-based obs. They are slow in torch, but maybe that is just torch. At least now we can benchmark conv-based encoders vs. alternatives at the actual speed of the hardware without worrying if any of those 5 wrappers were just slow.

Nano Banano 2 does a great job at generating explanative infographics purely from source code

@SemiAnalysis_ It’s Micron’s own fault, one doesn’t build things in NY of all places. Who decided to go for one of the states with the nastiest bureaucracy? Build in Nevada.

Micron’s $100B megafab in NY is at risk of delay due to just 6 “concerned citizens” and their frivolous lawsuit. (1/10) 🧵

@gabriel1 Government money is going to have negative impact. It is the best way to keep researchers in cozy, state funded positions where they will never do anything important.

the problem with france putting $30m into ai isn't that it's a small amount, but that they think the government allocating money works it might as well have been 3 trillion and they would still fail, just let businesses do business