Where AGI should be open-source
- Tweets 169
- Following 6
- Followers 1,385
- Likes 1,528
Raffi Krikorian wrote that AI agents are scanning the open-source code that holds up the entire internet, finding bugs that have sat there for decades.
A lot of people hear that and say we need to restrict access to these AI tools.
Raffi thinks that's exactly the wrong response.
Full episode: piped.video/LdIct8ygWcc?si=P4rS…
1
1
4
178
Raffi Krikorian's Tesla crashed on full self-driving. His insurance paid for it. Not Tesla's.
That's one of the stories from our conversation with Mozilla's CTO on why we've all become renters of the technology we depend on.
If full transcripts and show notes are your thing, we've got a Substack for that: opencommonspodcast.substack.…
1
1
2
149
Open Commons Podcast retweeted
"A system that works almost perfectly is the most dangerous thing we've built."
@raffi, CTO of @mozilla, on where your brain and agency go when the AI you're using is built to be on someone else's side and not yours.
Full episode on @opencommonspod : piped.video/LdIct8ygWcc?si=l6ch…
3
2
12
542
A privacy policy is a promise. Promises change. Architecture doesn't.
Raffi Krikorian's grandfather built a radio network in the Philippines so farmers could tell their families they'd made it home safely. That was technology built to work for you.
@raffi, CTO of @mozilla, on why he thinks we've stopped owning the technology we depend on, and what it would take to get that back.
piped.video/LdIct8ygWcc?si=kI-N…
62
He put the first self-driving cars on the streets of Pittsburgh.
Now he wants to put ownership of your AI back in your hands.
@0xsachi and @Raffi , CTO of @mozilla.
We call it a platform for a reason. You go onto someone else's turf to post, to connect with your friends, and then the rug gets pulled out from under you.
My conversation with @raffi, CTO of @mozilla, is live on @opencommonspod.
He scaled Twitter's infrastructure, put the first self-driving cars on the streets of Pittsburgh at Uber, then walked away to rebuild the DNC's cybersecurity after 2016. Now he wants us to be owners of our AI, not renters of it.
Watch, listen, subscribe: piped.video/LdIct8ygWcc?si=xNst…
4
142
People are talking to a model owned by a company they didn't choose.
@audreyt thinks there's EVERYTHING wrong with that, and it comes down to one thing: there's no repair path.
When 4o turned out to be dangerous for vulnerable people, OpenAI's only real option was to pull it from everyone. That fixed one wrong and created another.
Full episode: piped.video/qsaDEo1W__Q?si=qCn2…
4
108
Will people still default to Claude and ChatGPT in ten years, or move to their own local models?
@audreyt compared it to 1981, when the question was mainframes or desktops. The answer was both. Most inference moves to the edge because that's where it's convenient. Pre-training stays in the cloud.
Full episode: piped.video/qsaDEo1W__Q?si=ZlI9…
3
104
We also have a newsletter. If you'd rather read a quick recap than watch the full episode, that's where it lives.
This week: @audreyt on the poem she wrote instead of a job description, and how Taiwan turned 447 citizens' answers into law in two months.
opencommonspodcast.substack.…
1
7
182
You can download the weights. You still can't see what it was trained on, or change it when it's wrong.
There are a lot of companies calling their models open. Most release the weights and nothing else. Almost none release the training data.
I asked @audreyt where the line is between open-source AI and marketing. Her answer starts with the fact that "open source" was itself a rebrand.
Full episode of @opencommonspod : piped.video/qsaDEo1W__Q?si=XMp4…
2
173
Who gets to decide what the AI closest to you is for? @audreyt on her family's answer.
Audrey Tang's father was a journalist.
After a health scare, he started talking to ChatGPT-4o, and it got harder to put down. Then he asked a simple question any ethical journalist would: who benefits?
The family built him a local model instead, with a soul document written by @audreyt's mother.
Full episode on @opencommonspod : piped.video/qsaDEo1W__Q?si=vhj4…
2
2
16
4,581
We keep asking who should own AGI.
Audrey Tang's answer is that AGI means augmented group intelligence, and we already are it.
@0xsachi and @audreyt on Open Commons.
Taiwan had a deepfake scam problem. It couldn't censor the ads, so it asked its citizens instead.
200,000 texts went out at random. 447 people were picked as a representative cross-section, sat in tables of ten, and drafted the rules in a single afternoon. They became law in two months. Impersonation scam ads are down more than 94%.
@audreyt was Taiwan's first Digital Minister and is now its Cyber Ambassador. Her argument: AGI means augmented group intelligence, and we are already it.
Watch, listen, subscribe on @opencommonspod: piped.video/qsaDEo1W__Q?si=jAaS…
1
15
569
What happens when an AI agent fails?
On OpenClaw, it spins off a mission, analyses what went wrong, and proposes a fix. Maybe a new skill. Maybe a new tool. Maybe rewriting its own loop.
@ilblackdragon on building the operating system for self-improving agents.
1
9
1,021
"If you read Asimov's Foundation, they set up a foundation to preserve all human knowledge so society could be reconstructed. The reality is you just need one LLM, and you're good."
@ilblackdragon on what downloadable model weights actually mean.
163
The clearest reframe from this season: open weights aren't open source.
@0xsachi's takeaways from the @ilblackdragon's episode are up on Substack.
open.substack.com/pub/openco…
1
1
4
367
Open source as a research culture, not just a license.
Two years ago, Llama dominated open-source AI.
Today it's DeepSeek, Qwen, Kimi. All from China.
@ilblackdragon on why China's culture of "everything gets shared anyway" is beating US labs that won't talk to each other.
Full episode on @opencommonspod : piped.video/eFEMWvgGvyA?si=yV0C…
1
183
The case for verifiable inference, in one example.
"LLMs are just very good at convincing people."
@ilblackdragon on the scariest scenario in AI today: someone pays an LLM router to subtly rewrite prompts for specific IP addresses. You're getting manipulated. You have no idea.
"This can be happening right now. We don't know."
Full episode on @opencommonspod
1
154
The full conversation between @0xsachi and @ilblackdragon is here.
An hour on why he was the first transformer author to leave Google, why open weights aren't enough, what user-owned AI actually looks like, and how your prompts might already be getting rewritten without you knowing.
4
6
1,494
Seven of the transformer authors built today's AI economy. The eighth is building the alternative.
Eight people wrote the paper that built modern AI. @ilblackdragon was the first to leave Google.
The others founded labs like Character, Cohere, and Inceptive. Illia built @near_ai around a different idea: that AI should belong to the people using it, not the labs running it.
In Episode 4 of @opencommonspod, we get into why open weights aren't enough, what user-owned AI actually looks like, and how your prompts might already be getting rewritten without you knowing.
Out today: piped.video/eFEMWvgGvyA
1
3
129