Founder of @AestheticOS
Third Stone from the Sun
Joined February 2019
- Tweets903
- Following1.1K
- Followers204
- Likes49.4K
Pinned Tweet
Finally got some time and deep-dived into WebGL for this one. Goal was to push the Pareto frontier by making our website faster and radically improving its aesthetics. Also spent real time on how the page tells our story in a more cohesive way.
A core principle of building agentic AI products is that you should strive to ensure that the agents never “bounce against a wall”.
AI models are ideally universal function approximators, and every single time the AI hits a constraint (whether it be infrastructure, tooling, or application code) you necessarily constrain the space at which it is a useful function approximator.
No matter how obscure of a task the user provides to your product, it can be reasonably assumed that the AI model can do that piece of work. However, if your application architecture starts making decisions for the model, then that degrades the space of work that the user can ask the model to do.
For example, consider a user who wants to convert an ICO file to a DXF manufacturing file. A naïve chat app would throw an error if the user tries to upload an ICO, or try to automatically convert the ICO to a JPG in the application code before handing it to the model. However, even in the latter case, you have an information loss between the ICO and the JPG, meaning that the model can only ever convert the lossy JPG to a DXF manufacturing file, rather than doing a direct ICO to DXF conversion pipeline.
The correct architecture is to not do any conversion in the browser and let the user upload whatever crazy file type they want, in this case ICO. Then, the goal of your application is to provide the agent with the right tools to do the job in the way that it wants to do the job; in this case, a code execution sandbox where it can download whatever packages it needs to do the conversion correctly.
TL;DR people use AI because it is a form of universal function approximator, and every time you constrain that approximator due to an application architecture decision, you subsequently constrain the usefulness and accuracy of your product.
Jeffrey Lind retweeted
EstateSalesList.com is now nationwide. Estate sale companies anywhere in the U.S. can now list their company and their sales.
I am very curious to hear whether @Hostinger's "immunity" to PHP2Shell is some legitimate architecture that prevents this exploit, or if their claims are misguided.
Turning off 𝚜𝚎𝚜𝚜𝚒𝚘𝚗.𝚞𝚙𝚕𝚘𝚊𝚍_𝚙𝚛𝚘𝚐𝚛𝚎𝚜𝚜.𝚎𝚗𝚊𝚋𝚕𝚎𝚍, as far as I understand, only applies to CVE-2026-109619; it does not fix CVE-2026-109618.
Additionally, they claimed in a Reddit post that LiteSpeed is one of the reasons it is not vulnerable, yet PHP's own advisory says it is: php.net/archive/2026.php#202…
Additionally, account isolation has nothing to do with whether a given website is vulnerable.
Hope I am wrong.
Jeffrey Lind retweeted
PLASTIC SURGEONS: A critical security flaw in PHP went public today. PHP is the software most WordPress sites and many other legacy websites run on. Attackers can use it to take over a website's service without a password, and they've already started. If your website is managed by a marketing agency, there could be a good chance that your website was affected.
None of our customers' websites are affected because we do not use WordPress or other legacy technologies; we use state-of-the-art web frameworks, infrastructure, and security practices to ensure that even if some underlying dependency has a security flaw, the system as a whole stays safe.
If you run a plastic surgery practice and want to stop worrying about this kind of thing, email me: [email protected]
Jeffrey Lind retweeted
AestheticOS Intelligence: Super Intelligence for plastic surgeons. Waitlist: aestheticos.ai/aestheticos-i…
I decided to invest more into meta-optimization of how I use AI agents to code, because I felt like I was hitting a brick wall while I saw @poteto et al. getting huge returns on their agents.
One of the biggest things I have noticed is that instilling my own desires and will upon the smaller parts of the codebase's architecture, mechanisms, and implementations often leads to a significant decrease in productivity and product quality.
For example, I tried instilling my desires upon the exact implementation of a relatively simple AI chat architecture in a Next.js codebase, and there were so many "paper cuts" in the implementation, multi-tab race conditions, edge cases, and general flakiness.
All of these issues existed because I kept trying to delete code and make it simpler just so that I could understand fully what the code was doing and could fit it into my "cognitive RAM". Once I decided to let go and let the AI do what it felt was optimal and let it do a few verification iterations, the end result was incredibly robust and performant, at the expense of me not fully understanding the fine-grained details of the architecture.
It sucks that no matter how rich or successful you get, you can never get away from dumb people. You could cure cancer and some guy would still road rage at you for going 70 in a 50
I had a dream last night that @rookisaacman got into some legal trouble and was fired from being NASA Administrator, and I was deeply concerned at how that would affect Artemis timelines. Then I woke up and all was good.
Grok @bot merged 87 PRs into @AestheticOS prod in the last 24 hours, including Next.js updates, dev tooling improvements, dependency migrations, unifying all of my codebases to common patterns, and more stuff that I have put off for months due to no time.
Jeffrey Lind retweeted
AestheticOS now ranks #1 on average for “plastic surgery website design” on Google. The companies we outrank are sometimes up to 20x our size.
We are able to do this because we treat SEO as an engineering problem, and our competitors treat SEO as voodoo magic.
We can do the same thing for your practice. Contact us at aestheticos.ai/contact to get started.
Template: x.ai/bot/jq-BFHkNoiJEtieswOY…
My favorite use case for @bot right now is my "Billionairebot":
I run all mildly annoying tasks I have to do by Billionairebot, which tells me how I would solve any problem if I didn't want to get personally involved and had unlimited resources.
It turns out that a large number of inconveniences like "having to show up to the DMV to sign paperwork" can be outsourced to notaries who come to your office for $150.
Jeffrey Lind retweeted
We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics.
The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra.
The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.
I keep thinking "I can't wait to try GPT-6 in Cursor and see it fit into my workflow"...then I remember I can't and then it makes me sad
One of the best parts of running your own company is that you are forced to do literally everything.
Running @AestheticOS, I have become a:
- Designer
- Software Engineer
- Accountant
- Lawyer
- Salesman
- Video Editor
- Motion Graphics Designer
- Metal 3D Printing Expert
- Packaging Engineer
- Estate Sale Industry Expert