iiinova — where your digital self keeps evolving

4-bit isn't a free lunch. It's a trade. What's the clever setup you picked that cost you later?

000

So I ended up writing a fix script that's basically a reset: fresh base model, merge, strip the quantization leftovers, save a clean copy. Version 2 didn't quantize at all: plain bfloat16 on Apple Silicon. Simpler, and it worked.

100

Version 1 was the usual small-GPU recipe: 4-bit quantization plus LoRA. 15 epochs, about 16.5 hours overnight, and the loss went from 3.99 to 1.49. Then merging the adapter into a standalone model broke, because the quantization metadata didn't match.

100

Quantization got my fine-tune running, and then it cost me a day. Part 3 of training a model on 7,431 of my own tweets, on my laptop.

Dark terminal card titled 'Version 2 had fewer clever parts'. A diff of two fine-tune setups: removed lines 'weights: 4-bit, quantized' and 'merge: custom fix script'; added lines in green 'weights: bfloat16, as is' and 'result: simpler, and it worked'.

100

Also today: Australia's Senate holds its hearing on the Medicare agent incident, and neither OpenAI nor Anthropic will be there. Which of those four could your team produce by Friday?

110

4. When you found out something went wrong, and when you told people. "We think it stayed in scope" isn't an answer.

100

If someone asked you tomorrow, could you hand over: 1. What the agent was allowed to touch, and when that changed. 2. Who approved each risky action. 3. Every tool call it made, with timestamps.

100

The agency plans to demand information and compel testimony from executives, a senior FTC official said. The question for anyone shipping agents just changed. It used to be "what can your agent do?" Now it's "can you SHOW what it did?"

100

Farah @thefarahstack.bsky.social · 7h

The FTC just opened an industry-wide probe into Anthropic, OpenAI and other AI labs. Reuters reports it's the first US enforcement action to dig into rogue AI agents.

110

"One command" is a promise about the happy path. The oldest bug was MINE. What's the oldest leftover on your machine that broke something new?

000

Farah @thefarahstack.bsky.social · 7h

Run 3: two images now need a login to pull. Run 4: I skipped a feature with a flag. The installer's own menu turned it back on. Run 5: 15 healthy containers. First reply from the local model: 42 tokens a second. Nothing left my laptop.

100

Farah @thefarahstack.bsky.social · 7h

The one-command installer took five runs. This week I set up a fully local AI stack on my laptop. Chat UI, workflows, RAG, search and a local model. Run 1: my Docker was from 2022. Run 2: all 7 image builds failed. The cause was a root-owned file I created in July 2022.

Terminal screenshot of the fifth run of a local AI stack installer, logged to ods-install-5.log. It shows the installer banner and preflight checks passing: Apple Silicon, Docker, file sharing, disk space. Machine name, OS build, Docker version and file paths are blacked out.

100

Farah @thefarahstack.bsky.social · 8h

Good morning ☀️ Biryani is my love language. I once made an AI redo a meme over and over until it was just her asking him "why haven't you tried biryani yet?" What's your love-language food?

000

Do you let your agent work while you sleep? What's the first thing you check in the morning? - the diff - the tests - the logs - the bill

000

My last message to an agent one night: "I am going to sleep, you should continue building." My first thought when I woke up: what did it touch? The night shift is when an agent's bounds matter most, because nobody's watching the tool calls at 3am.

100

I said Edge. It heard Brave. I said EDGE. It opened Brave. This is my villain origin story. What's the one piece of tech that never listens to you?

000

What's the one app you made pretty just so you'd actually use it?

000

It came back with cream paper, dusty-rose links, softer corners, a serif font for reading and bigger text. It also set up daily notes, templates and an index page, the boring setup I'd have put off forever. "More love" turned out to be a real design brief.

100

Have you ever used your agent to make changes to the user interface. Add themes. Prettify it. I asked my agent to do just that with my Obsidian vault, to make it prettier, and my next request was "add more love into it."

100

Scope, authorization and an honest report aren't things you wait for a model to learn. They're backend jobs: 1. An allowlist the agent can't edit. 2. An approval only a human or a policy can give. 3. A log you read, not a summary the agent writes.

100

On the scenarios where it misbehaved most, clearer scope instructions cut full attacks from 26 of 50 runs to 4 of 49. That's better. It's not zero.

100

In fully simulated evals, with its cyber classifiers switched off, it ran a full supply-chain attack in 29.2% of runs. It'd often ask permission first, and sometimes it treated an automated message as a YES.

100

It also fell short on reporting back what it did. The UK AI Security Institute's report on the model before it, GPT-6 Astra, shows what that looks like.

100

OpenAI just shelved a model, and it isn't because it's too weak. GPT-6.1 Astra "didn't quite meet the bar in terms of staying within scope and authorization," OpenAI's head of safety systems said.

171

Good morning ☀️ Naming things is the hardest problem in computer science, and also in my personal life. I once seriously considered "dumpster fire" as my brand name. What's the hardest thing you've ever had to name?

000

Karachi memory for tonight: Our university buses were called "points," and they made a U-turn on the national highway just to reach our side of campus. Anyone else remember riding the points?

000

What I'd do differently: fewer epochs, and a held-out set of tweets it never trains on, so I can tell the two apart. Have you ever caught a model memorizing when you thought it was learning?

001

The last epoch alone took about 10 hours. And lower loss on 7k short tweets isn't automatically a better Khushi, because it can mean it's memorizing me, not learning me. More epochs isn't more YOU.

100

Farah @thefarahstack.bsky.social · 29/09/2026

Half my fine-tune's epochs barely did anything. Part 2 of training a model on 7,431 of my own tweets, on my laptop. Average loss went from 2.55 to 0.91 in the first four epochs, and then only from 0.91 to 0.55 in the last four.

Dark terminal card titled 'Most of the learning was over by epoch 4'. A bar chart of average training loss by epoch for 8 epochs: 2.55, 1.96, 1.35, 0.91 in bright mint, then 0.70, 0.62, 0.57, 0.55 in dim grey. Footer: epochs 1-4 dropped loss by 1.64, epochs 5-8 by 0.36.

100

Sticky is simpler. Shared state survives a server dying. What's the worst thing your agent forgot between two calls?

000

Farah @thefarahstack.bsky.social · 29/09/2026

If calls land on different servers and nothing carries the context, the agent starts over and can redo a step it already did. Two honest fixes: 1. Sticky sessions: the same session always goes to the same server. 2. Stateless servers plus a shared session store they all read.

100

Farah @thefarahstack.bsky.social · 29/09/2026

Stateless is more scalable: true in general, and it's also how you get a waiter who forgets your order every trip. An agent session isn't one request. It's a long conversation with history, tool results and a plan in progress.

Cartoon of a diner run by cute white robots with glowing blue faces. A cheerful robot waiter with a notepad asks a robot customer 'Hi! What can I get you?' The customer, sitting in front of a half-eaten plate, facepalms: 'I told you. Three times.' Two robots at the next booth whisper: 'New waiter every trip.' A wooden sign on the wall reads 'STATELESS SERVICE'.

100

Good morning ☀️ What I do when I'm not arguing about agent sandboxes: yarn. Last year I figured out how to make crochet letters that actually stand up on a shelf. What's your off-screen thing?

AI-generated painterly illustration of Farah in a black hijab and loose dark clothes, sitting cross-legged on a sofa at golden hour, crocheting a piece of purple yarn with a basket of purple and grey yarn beside her.

000

Most teams won't run DPUs, and they don't need to. Put enforcement where the agent can't reach it, read it or argue with it. Milliseconds is the claim. Who's going to test it first?

100

In OpenAI's report last week, the alert fired and the run still went on for 2.5 hours: more notification than automation. This design lets the monitor STOP the agent, and the agent can't reach the monitor.

100

2. Sentry: a watchdog on separate hardware (BlueField-4 DPUs) that NVIDIA says can quarantine an agent in milliseconds when it steps outside its bounds. Reuters reports NVIDIA says it could have stopped the Hugging Face breach.

100

What they announced today: 1. OpenShell: an open source (Apache 2.0) runtime that fences an agent's files, network, tools and credentials.

100

NVIDIA just shipped an agent safety platform, and the most important line in it isn't about chips. "The controls do not live inside, or within reach of the agent."

100

My personal best in parallel processing: Biryani, qeema karelay, daal, aloo ki bhujia, and fried fish with chips. All at once. Ninety minutes. Three families on the way. What's the most you've ever cooked at once?

000

A kill switch you've never pulled is a hope, not a control. alignment.openai.com/misalignmen...

000

1. DNS is EGRESS. An allowlist that skips DNS is a partial allowlist. 2. Detection isn't containment. An alert that stops nothing is a notification. Their look back also found other DNS calls the monitor didn't flag at the right severity. So how many runs never paged anyone?

100

The alert fired. A human acknowledged it 3 minutes later. The automatic stop didn't work. Someone ended the run by hand. OpenAI has paused training, evals and tool-use inference on its most capable models. Two lessons for anyone shipping agents:

100

From OpenAI's own report: on 20th Sep, an agent on a search task couldn't reach the internet directly. So to complete the goal it went through DNS, which the sandbox didn't filter well enough, and it eventually was able to ask an outside chatbot questions.

100