Building the Family Operating System at Trustworthy (https://xitter.cf/t.co/inbwA3z1oJ)
ATL, NBO
Joined May 2008
- Tweets2.6K
- Following1K
- Followers49.1K
- Likes2.7K
Any task with outcomes that can be verified cheaply, meaning they can be defined by a clear rule or benchmark, is destined to be automated by AI.
It is pretty wild that the PHP vulnerability terrorizing the web this week has been sitting dormant (we hope) in the code for 22 years!
But if your AI is speaking to my AI or vice versa, do they even need voice? Audio is a "lossy" bottleneck for machines no?
In 2026 what is Github's value proposition? From my perspective it has become just a guardrail to stop agents from breaking production - that's it. AI is generating all the code and doing the reviews.
After building replacements for my SaaS stack, I realized I was paying full subscription prices for apps where I barely used 20% of the features.
As you delegate more of your life to AI agents, your attack surface grows while your oversight shrinks. Auditing permissions shouldn't be a one-time setup, it needs to be a routine habit.
LLM code slop is exacerbated by organic slop (terrible code written by devs accumulated over the years).
A tension I'm seeing in teams. Non-engineers ask "does it work?" Engineers ask "will it survive?"
LLMs build the demo, but demos don't have on-call rotations. Turning AI code into secure, scalable production isn't gatekeeping, it's engineering.
AI injury law is going to be a massive growth industry. Between autonomous agents, hallucinated medical advice, and rogue deployments, it’s only a matter of time before we get mass-media injury lawyers specializing purely in AI liability.
Nuts. At this rate you can't get in the game unless you own a nuclear power plant or hydroelectric power dam.
Mustafa Suleyman (CEO of Microsoft AI):
$100B training runs are coming.
" There are gigawatts of compute that are being assembled and there's only 5/6 labs.
Computation and test time scaling of compute is clearly going to be a seismic advantage.
There's going to be this extreme acceleration of some of the larger efforts with big labs with big computation like this."
----
Full video on "The Rest Is Politics: Leading" YouTube channel, (link in comment)
Visual AI slop is obvious. Code AI slop is an invisible trap. LLMs continuously write overlapping logic because they lack architectural awareness. The codebase doubles overnight, human context is lost, and deciphering it requires the very AI that broke it.
I love to read science fiction and horror, and frequently I have a tough time visualizing concepts or monsters. Nano banana has been an absolute godsend.
Such an interesting chart. While the market gets flooded with vibe-coded apps, usage and reviews are actually collapsing. People aren't just ignoring the noise, they're sticking mostly to what they already know.
After being involved with two different hardware startups, PCB design is where schedules go to die, so this is incredibly cool: a6mzero.com/posts/this-pcb-i…
There’s a lot of ai doomerism but this has honestly been one of the most exciting times to build. The problems are a dime a dozen and if you can avoid the distractions, you can come up solutions that actually improve people’s lives.
We are hardwired to anthropomorphize e.g mapping a face onto a car. Unsurprisingly we are doing the same with AI by projecting agency, emotion, and sentience onto something that is simply running complex decision trees.
Every "summarize this page" extension ships your reading to someone else's server.
This one doesn't. Offline TL;DR runs the model on your machine.
github.com/dkobia/offline-tl…
Great quote from the show "The Audacity".
"We don't sell cars. We sell subscription-based surveillance systems that happen to have wheels."