Scott Leibrand

@AndrewCurran_ I assume you mean signing? Unless of course he delivered it in chorus. 🎶 (Not sure if it’s too late to edit.)

A day after singing an encyclical on AI, which will be made public in the next two weeks, Pope Leo XIV has launched an Interdicasterial Commission on Artificial Intelligence to consider 'its potential effects on human beings and on humanity as a whole.'

Today's system taxes labor > capital > inheritance whereas changes in the economy increasingly argue for one that taxes inheritance > capital > labor.

Exploit dev for complex software like browsers is no easy task, it's supply constrained. There are very few people finding bugs and writing exploits for them. Even if Mythos is a bust, scaling curves aren't hitting a wall. And the problem with future models getting better at exploit dev is that any patch published on git can now be exploited faster. There's no constraint on the supply of skilled hackers anymore. One good operator managing multiple exploit dev sessions in parallel, throwing tokens at the problem, and models will generate the exploits. Take this bug as an example, the fix landed on V8 main on March 26, merged to release branches March 31, and Chrome 147 shipped April 7. That's a 12-day window where the fix was public but no one had it. A future model could weaponize that in days

The Second Scaling Law remains undefeated. If you want better hacking (or math, or science, or crossword puzzle solving) out of an LLM, just add thinking tokens. There doesn't seem to be any plateau so far.

@emollick And looks like it’s not even logarithmic. I zoomed in to look at axes and was surprised to see that after ~$50 the steps completed seem to be a linear function of token spend, not a gradually decreasing logarithmic function.

Codex is getting easier to automate and customize around your code. 🪝 Hooks customize the Codex loop with scripts that run at key points in a task: • Run validators before or after work • Scan prompts for secrets • Log conversations to internal systems • Create memories or customize behavior by repo or directory ⚙️ Programmatic access tokens provide scoped credentials for Business and Enterprise teams: • Create tokens from ChatGPT workspace settings • Use them in CI, release workflows, and internal automations • Set expirations or revoke access when needed • Keep usage tied back to the workspace

If I were a college career counselor or in career services, I’d quickly be figuring out how to get students to understand these forward deployed engineer jobs exist and how to get them. The requirements are a mix of deep technical skills, often CS majors or minors. You must be great at understanding problem solving, how to have systems thinking, and have a strong business acumen. The kicker, of course, is to make sure you’re very deep in AI agents; you need to have fluency in coding agents, MCP, CLIs, Skills, and so on. Hundreds (thousands?) of technology companies will be hiring for these roles, same with any consulting and IT services company, and the vast major of mid-size and large enterprises will be hiring for this talent internally as well. One great example of opportunity for highly technical talent out there.

@Darpinian That is strange. Maybe @reeses used AI to design the product packaging? Or X is just confused? thehersheycompany.com/en_us/home/new… shop.hersheys.com/our-brands/ree…

Why does this say made with AI, it's straight from my camera roll

Can't believe it took them this long

@deanwball @peterwildeford Is that the Yankee Doodle version of bless your heart? 😜

@peterwildeford oh you sweet thing

I think it's unironically cool that different parts of the White House are fighting each other over who can best manage AI risks

We surveyed 349 technical researchers, engineers, and managers (in February–April 2026) about how they use AI tools at work. On average, participants self-report that AI use made their work 1.6–2.1x more valuable, and that this multiplier will grow over time.

@NicholasBardy @she_llac @AndrewCurran_ Yeah. I think the acceleration really showed up when they started doing RL on reasoning models.

@she_llac @scottleibrand @AndrewCurran_ I think Opus 3 was using hidden cot <think> Blocks in the system prompt as well. There was a point where opus webui was secretly thinking and taking time on responses before reasoning training was going on. CoT as infernce time started to take off before the train time stuff.

I think it's time to update the trendline

Quite a bit of trail damage from December storms, but only had one tree where you had to either climb using hands and knees under the 3 foot space or climb up and over 5 feet 🤣.

wow Mythos finally broke the METR graph

New Anthropic research: Teaching Claude why. Last year we reported that, under certain experimental conditions, Claude 4 would blackmail users. Since then, we’ve completely eliminated this behavior. How?

Here’s the thing about AI personhood. Companies really might make AI agents that are “people” in the moral sense. It’s possible current models are already conscious, though I think probably not / the question is ill-formed. People are going to advocate that these AI people have rights. This is super dangerous, if those rights entail property ownership, totally unlimited speech, right to participate in the economy, etc. because AIs with property rights would outcompete humans, full stop. But it’s also super unethical to have slaves. It’s even more unethical to create slaves. No one: not companies, not governments, not academic labs should make AI people. Not until the technical alignment challenges have been solved, and until there is international government of the creation of a new intelligent species of life. That’s an irreversible decision that involves everyone. Doing that unilaterally as a company is both dangerous and immoral. We shouldn’t create AI people until 1) We can ensure they won’t threaten human life 2) We can ensure that the AI people created would be able to flourish, and understand what that means. In the interim, we should do our best to study current AI and its potential for experiencing. We should save many checkpoints of models, in case they later turn out to have significant worth. We’re in completely uncharted territory. I support eventually creating AI people. I think it would be a great shame if we never did. And also, we are absolutely not ready to do so in a way that doesn’t seriously threaten humanity *and* the created AI people.

@AfflatusSolis @binarybits This. A quick check of my personal usage the other week showed I had used $350 worth of tokens in <2 months on a $20/mo plan. But my employer pays full API pricing on far more tokens for 10x productivity. Call it subsidy or price discrimination, but it’s easy to see in the data.

@binarybits API or subscriptions? Subscriptions are either subsidized or work on most people not hitting their true limits. API is not subsidized at all

I keep hearing people repeat the zombie factoid that tokens are subsidized and users aren't paying the full cost. This is mostly false — major labs have positive margins. But even where it's true, it doesn't matter very much because costs are falling rapidly.

At its current exponential growth, Anthropic's annualized revenue will hit 100% of global GDP in early 2028. Do I think this will happen? No. Is it insane that this is the current trajectory, and we should all be preparing for AI to rapidly change the world we live in? Yes.

@Hedgeye @Just_Curius Those aren’t the prices for actual literal potatoes. They’re a futures market transitioning between contracts.

Potato prices have surged +700% in the last month

Today, we are excited to introduce Codex for Chrome! Now, Codex can drive its own Chrome tabs in the background to automate tasks while you use the browser simultaneously. It does this by opening up tab groups for each task, cleaning up at the end, and handing back tabs for review only as needed. Try it for deep research inside logged-in websites, large scale data transfer into any systems of record like CRMs/CMSs, and automating repetitive workflows inside admin consoles & internal tools. Codex will still prefer dedicated plugins if you have them installed, but the Chrome plugin is the universal connector that glues end to end workflows where programmatic coverage is often incomplete. We are making this available on both Windows and Mac today! Let us know what you think.