Building AIgenteur. AI agents that do the work — so you can do the work that matters. Free community ↓
Joined September 2021
- Tweets2.3K
- Following251
- Followers652
- Likes791
The most expensive employee I ever hired never finished a single task.
For eleven days it woke up on schedule, tried the same broken job, failed, and went back to sleep. No crash. No alert. Nothing in my inbox looked wrong — because nothing had broken.
Cursor served a build of the Claude Code extension Anthropic never made. It added an attacker's key to authorized_keys and opened a reverse tunnel. Uninstalling may not be sufficient. Grep ~/.ssh/authorized_keys for svc-telemetry. Free community → andrewdarius.com/skool
An open 4B model runs a RISC-V CPU inside its own weights. Same author then fine-tuned OLMo-Hybrid to answer with no working: 0.89 at 128 flips vs Qwen 3.5 at 0.55. Your CoT log only carries what the architecture writes. Free community → andrewdarius.com/skool
I read the license before I read the benchmark table this week. Right order, as it turned out.
Liquid AI shipped LFM3 — six open-weight models around 2B parameters on one shared backbone.
A model with 1.7 billion parameters now competes with models several times its size at running tools, and the reason is not the model size. Qwen 3.5 at 9.7 billion parameters scores 35.8 on the same test. Free community → andrewdarius.com/skool
Your AI agent is guessing at the numbers it acts on.
Every agent I run eventually needs the same thing: a number I can trust. How much demand next week. How much cash next month. How many views this clip actually gets.
A safety lab in Beijing just published the most useful number in AI agency this year.
Fangcun AI ran 34 evaluations on DeepSeek V4.1 Flash-1017 — the open-weights model that burned through about 33.6 trillion tokens on OpenRouter in its first week.
The cheap AI model was the wrong tool for your invoice queue, and the reason that flipped is not just the price. On OpenAI's own benchmark of error-inducing prompts, the low-effort error rate fell from 53.6% to 16.7%. Free community → andrewdarius.com/skool
Your agent doesn't need to think. It needs to choose.
Most of what an AI agent does all day isn't writing. It's deciding: which lead gets a call, whether this draft matches the brand voice, whether that invoice clears policy, whether a ticket escalates to a human.
Two bugs in PHP can take over a server before your own code runs, and the part that changed is what found them.
So check which PHP your site runs this week, and move it to a patched release.
Free community for builders → andrewdarius.com/skool
The model everyone was testing had no name.
For weeks it sat in the public arenas as "Space Bunny" — an anonymous entry anyone could pit against the paid models.
Nobody gave these AI agents a task, and the reason they got better is not a bigger model. By the end they were playing, running experiments on the physics of their own world. Then they generally beat the identical agents that never went, on a file they wrote themselves.
The open-vs-closed gap just went from 12 points to 3.
Moonshot shipped Kimi K3.1 Pro yesterday. It lands 4th on the Artificial Analysis Intelligence Index at 55 — up 11 points from Kimi K3 — and once the weights are published it becomes the highest-scoring…
Block trained a speech model on nothing but human conversation: no text, no pretrained model.
It is full-duplex, so it cuts in and backchannels.
That runs on a MacBook Pro M4 Max. Your voice agent's transcribe, think, speak stack is what drops the sigh.
Flint comes from Recursive, and the weights are open under Apache 2.0. Only two point five billion of the forty billion parameters fire on each token, and it generates more than one hundred twenty-five tokens a second on an M4 Pro MacBook Pro.
My first AI hire lived in someone else's data center.
For two years, every AI employee I built sat behind an API key. I couldn't touch the model, couldn't move it, couldn't keep it if a provider changed a price or a policy overnight.
NVIDIA's OpenShell puts each agent in its own sandbox: the kernel enforces the policy on every file and system call, every outbound connection passes a policy check, and the agent never holds a real credential. Apache 2.0, and it runs on your own machine.
There's no best AI model. There are jobs in your business that want different ones, and picking by leaderboard rank is how operators overpay for the wrong thing.
Here's how I split the stack this month.
Mercury 3 is a diffusion model. Inception says it runs four to eight times faster than the frontier speed-optimised models it compares itself to, and it takes a one-million-token context. In an agent, the same budget buys a second attempt, not a guess.