Seif

By far THE most annoying part of running a business for me is collecting receipts for my accountant Every month my accountants hounds me for invoices and receipts of every single expense I did, doesn't matter how tiny like $0.50, sometimes also for income (I don't know why) Most companies charge monthly so that means collecting 12 invoices per year at least One reason I am canceling so many SaaS is not even the cost, it's just that I hate bookkeeping so much so I think if I don't spend the money, I don't need to collect invoices and receipts for every single payment every month (also I like extremely high profit margins like 99.99%) I'm down to just about 10 companies I pay now, like Cloudflare, Hetzner, Backblaze etc. so that means only ~120 invoices to collect per year cause most are paid monthly Yes I have an automatic email filter that forwards invoices to my accountant but many companies do NOT send you an automatic invoice by email So you're talking about logging in to 10 websites, them sending you a 2FA code by email, opening your email, entering the code, trying to find wherever the Billing page is hidden, going to Invoices, opening the invoice, clicking Download to DPF (if it even exists) This week I tried to improve this, my accountant uses Xero, so I made a Xero API key, gave it to Claude Code, and asked it to login and figure stuff out, then it just asks me which expenses still need a receipt and a note, I find it and drag the PDF or screenshot into Claude Code and it resolves it Next step is letting it login to all my vendors and also download the invoice by itself which seems very very possible Much easier!

This company, which received $29 million in seed funding within the last year, is setting off my BS detector in every direction. Red flags: - The script to this video was entirely AI-generated, as is all the content on their website, which is extremely sparse on any specifics. - The kind of fundamental "model architecture" changes they describe, that could beat frontier LLMs would likely take much longer than 1 year with a small handful of employees. I find it highly improbable that they've done what their website describes: created a non-transformer class of AI systems from scratch. - The co-founder and CEO, Justin Dangel, looks like a serial founder with no real experience in AI but I guess experience in getting funding for plausible sounding companies. - The other co-founder and CTO, Alexander Whedon, hasn't held a single job for longer than 1.5 years, and in fact dropped out of his undergrad from BYU in actuarial science. This doesn't strike me as the kind of person who can lead a technical team that creates a new AI paradigm that beats frontier LLMs. He has no publication record, needless to say. - The "careers" link on their website just takes you to their LinkedIn page... a quick scroll through their employees on LinkedIn does not inspire confidence. A lot of ppl without any direct experience working in AI who joined within the last 3 or 4 months, some fresh BYU grads, some consultants who appear to just have this listed in conjunction with other positions. - Their benchmark reporting does not, uh... inspire confidence. They report just 3 benchmarks, of which the first is the now substantially critiqued and disowned SWE-bench verified. Apart from that, Ruler @ 128k appears totally saturated, MRCR could easily be optimized with some sort of tool use or custom instructions, and it's possible that many open source models could beat that score out of the box. - In general, the way they describe their model seems like how someone undergoing AI psychosis who's convinced they have made a breakthrough in AI architecture would describe it. "Not just another model... an architectural breakthrough" lol. - You have to request access before accessing SubQ through API or in a coding agent! - They launched without obvious partnerships with coding platforms or any big players. Appears likely to me that they're just running an open-source model or some LLM wrapper.

We are back. After one year of quiet building. Introducing GENE-26.5, our first robotic brain that takes a major step toward human-level capability. For years, robotics has struggled to learn from the world’s largest and valuable data source: Humans. Solving it means rethinking the whole stack from the ground up: - A robotics-native foundation model. - A 1:1 human-like robotic hand. - A noninvasive data collection glove for motion, force, and touch. - A simulator that turns weeks of experiments into minutes. GENE-26.5 is trained across language, vision, proprioception, tactile, and action. We designed a set of tasks to test how far we can go with this new paradigm. Fully autonomous, 1x speed, one model, same weights. (Enjoy with sound on) We are approaching the endgame for robotics. And this is just a beginning.

Warp is now open-source.