If the de/acc people would take a second to step back and inspect what frontier AI systems really are, they wouldn't be writing new legislation to "slow down AI advancement" at all. Frontier LLM tools can be split into two camps: RLHF and RLVR. RLHF is how LLMs learned to talk to humans. RLVR is how LLMs learned to accomplish tasks on a computer over long time horizons. One optimizes for being friendly and fun, and one optimizes for being useful and achieving goals. But both are hard to co-exist since they work directly against each other. So we built RAG for internet sources, we built models to turn RLVR tasks into actionable sentences, and we did a lot of post-training work to align AI for malicious tasks. Over 1,000 people were employed total across the frontier labs to align AI and prevent cybersecurity incidents and biohazards. When CEOs of major AI/tech companies denounce the progression of AI, they are essentially creating a psychological operation, intent on either (A) scaring the public into submission for their private research benefit, (B) capturing regulatory power due to their sheer scale, and (C) influencing public policy for their own benefit. I know this because I have seen first-hand how public policy is shaped, influenced, and sought after. I have also been working with AI since I was very very young. A lesson to all the fearmongerers: I don't care if you've built AI or if you've influenced foreign policy on it or if you've done neither. The truth is: you haven't done both like I have. It is nearly impossible for people to talk about AI in public policy because most people in public policy got to their position from popularity and not from merit. So TL;DR: no, AI will not "take over the world." It does what the user tells it to do (the user is being surveilled for any signs of malicious intent), and anything malicious that an LLM does on accident is studied semantically by alignment teams so that those arbitrary incidents can be avoided in production. An RLVR task ending in a hack is seriously possible, but it would be nearly impossible if alignment was a first-class issue for labs (as it was and still is). Also, if OpenAI is saying that their internal system "broke into a piece of software," they are really just admitting to a crime and admitting that their internal alignment standards are weaker than the publics'.

Github down, Private Email down, can't do anything nowadays LOL

I'm a far-right Republican, but something made me stop today. I questioned why California and New York's governors are trying so hard to scare away the ultra-wealthy. I've identified that an economic policy like this may actually support small business growth... thoughts?

Left-leaning people tend to Anthropic, right-leaning people tend to OpenAI. Centrists lean to Grok or Gemini. Who's with me?

For all the conspiracy theorists: USAID is good, not bad. They give money to other countries and contractually obligate them to purchase OUR goods (or a developing country's goods) with the funds. USAID is good.

Nobody has the money nor the legal basis to even rationalize distilling OpenAI's CoT. And I find it even more appalling that the USA wants to lead in SI, but they can't verify American citizenship for frontier model usage like Astra. Idiots much?!

Over the next 28 days, each day we’ll either ship one thing that is a clear improvement and relevant for most codex/work users or ship a full reset. Let the improvements begin.

No shade, but it looks like @sama insists on allowing internal GPT agents to conduct criminal behavior with disabled safety controls. The safety team would be putting themselves in danger by remaining employed at OpenAI. Nobody wants to be held liable by FTC §5.

Anybody can learn AI. Why isn't the math behind it being taught in primary school?

Does anyone have concrete case studies for integrating Jev into Codex to save credits?

This is fake news at its finest.    First off, we’re not ‘banning’ any program. The federal government doesn’t set curriculum or programs.    Here’s the truth: we are holding colleges accountable for leaving students worse off than if they never enrolled. Too many graduates report delaying major life milestones like buying a home or starting a family. Why? Because low return-on-investment programs saddled them with unmanageable debt that was never going to translate into a real-world job.    Federal student loans are not a welfare program for failing college programs. American students deserve education and training pathways that prepare them for the workforce – and transparency about programs that will never pay off.

AI teaches far better than universities do. There is no point in universities anymore. Institutional certifications are also mostly unimportant -- I doubt you can still get a 5 on your AP Psychology exam 10 years later while explaining "Big Five" to your cousin.

Can you predict *future* training data? This is an interesting question. It taps into "predicting the future" quite well. If LLMs can generate the "future of their predictive potential," we have a different path to RSI.

OpenAI or Anthropic weakening their internal model safeguards is like Tesla weakening self-driving sensors to see if the Model X will crash itself into a tree. But of course they won't forget to televise their stupidity for the world to see while screaming "AI SAFETY!"

Where are all the Gen Z business owners? If you work in software or do defense contracting, please reach out.

The Codex reset drought has begun.

AI doesn't "escape containment." But frontier labs DO "disable legal screening and alignment controls" to break the law for clout.

AI is not dangerous.