I work in artificial intelligence and in my free time I like to work on artificial intelligence and sometimes I like to talk about artificial intelligence
United States
Joined July 2025
- Tweets526
- Following821
- Followers208
- Likes5.6K
Pinned Tweet
Why are they playing philosopher in my coding machine
Nah. lol.
Remove your main @github account from your computer. Now.
Give your AI agent a separate GitHub account with minimal permissions per repo.
Let the agent create PRs, then use your main account to review and merge them.
Obvious advice, but essential in the age of AI agents.
Claude Projects had two weeks where they actually worked. Now the orchestrator is so lobotomized that you have to micromanage the subagents yourself anyways. The company deserves to fail
Bizarre how Anthropic gave us the best model of all time for long-running tasks and then gradually lobotomized it to the point where it can’t run for 10 minutes without asking for some unnecessary permission and stalling out
When did everything become about “automations” and “workflows”? This has never been the point, and this sort of language only started getting popular over the past 6-12 months. Why would we take superhuman capability and use it for a workflow that a human already does?
This is correct. The vast majority of enterprise workflows have not actually had agents applied to them in any meaningful way. The economy probably has this deployed < 1% right now.
At best you have a few automated systems in a company that do a single pass of an LLM to make a decision or provide feedback, but very little real “automation”. Some of this is a statement about the adoption rates of technology in enterprises, but the vast majority of it is really just due to the capability of AI itself.
That’s all changing. Now with computer use, long running agents, harness improvements and surrounding infrastructure, we’re finally at a point where we can start to apply agents at more complex workflows.
We will see self-improving systems that can take any particular workflow goal and improve the execution rates over a volume of instances. And with the emerging form factor of agent swarms, you can just choose to deploy any amount of compute at a problem and agents will off and do the work.
This won’t be about and users prompting agents to complete a specific task for them, but instead the organization reengineering the underlying workflows to take advantage of agents in their processes. This will have a massive impact in finance, life sciences, healthcare, and other information-heavy domains. Lots of opportunity to build and deploy these systems right now.
Workflow automation is going to share the fate of those elaborate guardrail systems we used to make for sonnet 3.7
Imagine a PhD at the University of Wyoming, -12 degrees outside as you’re heating the building with 90 A100s
Now that ability to scale technology is more important than showy faculty prestige, the flagship state schools are looking like great places to be. I bet the less financially comfortable top tier colleges will tank as the serious public schools rise
Code Monkey retweeted
Good list. If you already work in IT, here's the next step:
- A guest network is the beginner version (not trying to be condescending). Split IoT, TV and guest into their own zones behind your own router (not the ISP box), block everything by default, then open exceptions one at a time. Block those zones from reaching private ranges, and do it for IPv6 too, or traffic just walks around your zones.
- Know which of your exceptions can be faked. Allowing a device by MAC or source port is convenient, but not secure. Same with hiding your SSID, which NSA says adds no additional security.
- Test it from the attacker's side. Put a real client on guest and IoT and try to reach your LAN, then scan from outside. A service reporting active doesn't mean traffic goes where you think it does.
- Anything that depends on the VPN (and you SHOULD be using one, and no, NordVPN is not a valid option; buy Proton or Mullvad) goes in its own network namespace with its own wireguard tunnel, so if the tunnel drops, it loses connectivity instead of quietly leaking. Then actually drop the tunnel and check.
- Kill LLMNR and NBT-NS (and mDNS where you don't need it), and require SMB signing. Poisoning is only half the attack; relaying the creds is the other half.
- Router: no remote admin, UPnP off, WPS off (the PIN falls in 11k tries). I also keep management on key-only SSH bound to the LAN, and I re-check my settings after every firmware upgrade, because upgrades love quietly turning stuff back off.
- MFA isn't enough anymore. Admin and email accounts get passkeys or FIDO2 keys. SMS, TOTP and push can all still be phished.
- Every network change gets a rollback timer and an SSH session you keep open, so one bad firewall rule can't lock you out of your own router.
- LLMs are great for reviewing configs, but redact keys, passwords, public IPs and MACs before you paste anything into a hosted model, and verify what it tells you. It's not an audit tool by itself.
- Backups have to be offline or immutable, and if you've never tested a restore, you don't have backups. Alerting shouldn't run on the box it's watching either. If that box is the one that dies, nothing tells you.
Here's my own network for reference, anonymized:
There has been a surge of cyber attacks happening all over the world
I strongly recommend doing a full security audit on:
1. Your home WiFi network using one of the LLMs. At minimum get all your IoT devices on a guest network
2. Any code repositories you have deployed anywhere and even those you only run locally
3. Passkeys and MFA. Stop using that password from 8 years ago and move to a password manager like 1Password
All a pain to do but a tiny amount of time relative to cleaning up a potentially large mess later
The labs are A/B testing usage limits out in the open and people still think it’s ridiculous to suggest that they might be quantizing or conspiring
I’m starting to wonder if I was wrong, and opus is not exceptional at aesthetic judgement, but they’ve tricked me with an extremely detailed rlhf facade
Is there a bench that correlates with authentically novel idea generation?
Now is probably a good time to revisit those permissions you gave the coding agents without a second thought
“It added a lot of dead code and unnecessary complications”
Who cares? What are the actual consequences of this?
"I don't review code anymore"
Honestly, I'll never understand this.
I just had to prompt Opus 5.5 more than it was necessary to simply convert a dialog to a popover.
It left a lot of dead code and added unnecessary complications. I had to read the code multiple times and point out the issues.
Don't tell me it's my skill issue for such a banal change.