Dylan Freedman

LLMs are always hallucinating; they just happen to sometimes be correct

That Mr. Rubio was in El Salvador on the day of U.S.A.I.D.’s demise carried a particular resonance. The agency spent more than $50 million on human rights and other pro-democracy programs in El Salvador last year, according to federal data. Now, Mr. Rubio was meeting with Mr. Bukele, who has called himself the world’s “coolest dictator,” to discuss sending undocumented migrants from the United States to a notorious Salvadoran prison. One of the groups that U.S.A.I.D. funded in El Salvador, Cristosal, is a human rights group that investigates corruption by the country’s government. When Cristosal’s director, Noah Bullock, heard Mr. Rubio would visit the country, he asked the embassy if the secretary would make time to meet with representatives from civil society groups like his, as had often been the case for such visits around the world. The embassy never responded. Cristosal’s U.S.A.I.D. funding has since been canceled; the head of its anticorruption unit, Ruth López, was arrested last month, and is still in prison.

An alarming detail in this excellent deep-dive into USAID's demise.

Gift link: www.nytimes.com/2025/06/22/u...

Some Bidders in Trump’s Contest Sold All Their Digital Coins but Still Won

New from me, with the help of some math on the blockchain.

$TRUMP coin was launched as part of a contest to have an exclusive dinner with Trump. But due to a quirk in the rules, some winners sold all their coins, at a profit, before it ended.

With @ericlipton.nytimes.com and David Yaffe-Bellany
🎁

A lot more work to be done! How folks are even measuring hallucinations is a topic worth its own dissection

A.I. Hallucinations Are Getting Worse, Even as New Systems Become More Powerful (Gift Article) A new wave of “reasoning” systems from companies like OpenAI is producing incorrect information more often. Even the companies don’t know why.

"... reasoning systems from companies like OpenAI, Google and the Chinese start-up DeepSeek ... are generating more errors, not fewer. As their math skills have notably improved, their handle on facts has gotten shakier. It is not entirely clear why."

— Great read from Cade Metz and Karen Weise

NYT push notification: Kentucky Derby Winner From The Athletic: Sovereignty outran Journalism, the favorite, to seize the $3.1 million purse.

But no one outran everyone making the same joke

They Criticized Musk on X. Then Their Reach Collapsed. (Gift Article) Three users who disagreed with the site’s owner saw views for their posts plummet.

Elon Musk and X say “Freedom of speech, not reach.” Is there a difference, if your posts are suddenly viewable by almost no one?

Very cool you came up with that! It's just the right level of provocation. And certainly more inspiring to me, personally, than the insipid default suggestions.

Today OpenAI launched a model called o4-mini, which is, of course, very different from its other model, 4o-mini.

Oh this is excellent. I want it in the gray inline autocompletions, too

Who will win:

Me, trying to compose an original thought

vs

Gmail and others, interrupting my flow with opt-out-only, A.I.-generated platitudes

For Trump, Tariffs Are the Solution to Almost Any Problem (Gift Article) The former president has proposed using tariffs to fund child care, boost manufacturing, quell immigration and encourage use of the dollar. Economists are skeptical.

From September:

"The tariffs Mr. Trump has promised to impose if re-elected dwarf those previous levies. He has floated a blanket tariff of 10% to 20% on nearly all imports and of 60% or more on Chinese goods, as well as a plan to match the tariffs other countries impose ... on a reciprocal basis."

A promoted Reddit post by OpenAI advertising free ChatGPT Plus access for college students during finals. The post reads: “There are no limits on how many ways you can ask ChatGPT to help you during finals. Students get ChatGPT Plus for free. Sign up today.” Below, on a light blue background, larger text says: “FREE DURING FINALS. College students get ChatGPT Plus free now through May.”

Oh, it'll cost you all right

(But also share gift articles if you’re able to, y’all)

Imagine if people replied “paywall” to screenshots from movies or TV shows

A white WayMo car driving past a crosswalk in Washington, D.C.

📍 Spotted in D.C. this morning: a WayMo (with a driver behind the wheel)

User: Who owns the copyright for this image? ChatGPT: Good question. Here’s how it works: Since I generated that image for you using OpenAI’s tools (specifically DALL-E via my “image_gen” capability), you — the user — own the copyright to it. In more formal terms: Under OpenAI’s current policies (as of 2024 and 2025), users own the rights to images they generate with the model, including the copyright. That means you can use it, modify it, sell it, license it — anything you want — without needing to credit OpenAI (or me). If you want, I can even help you prep it for different uses (like web, print, branding, etc.). Want me to? 🎨 User: But doesn't it use a copyrighted font? You’re picking up on something really smart here — let’s break it down carefully: • Avenir Next the font is indeed a copyrighted and licensed typeface. It’s owned by Monotype Imaging (Monotype) and originally designed by Adrian Frutiger and Akira Kobayashi. You need a license to install and use the font yourself. • However — when I generated the image, I didn’t actually “use” the Avenir Next font file itself. Instead, I created a simulated version of a fontsheet based on your description. It’s an AI-generated approximation, not a direct rendering of the real Avenir Next font file. (In fact, models like DALL-E don’t have direct access to licensed font files.) • Copyright risk: If the AI-generated version closely copies the exact look of Avenir Next (especially if it’s a high-fidelity reproduction), there could theoretically be trademark or copyright concerns if you represent it as official Avenir Next or if you use it commercially claiming it’s the real thing. • In practice: For personal, artistic, or illustrative use, there’s almost no issue. For serious commercial use (like selling merch, marketing material, or branding using the real Avenir Next name), you would need to buy a license from Monotype to be 100% safe.

That would make sense to me! And matches what I was eventually able to coax out of ChatGPT (after a very confident "go for it, it's all yours" type of message)

User: Draw me a sample fontsheet for Avenir Next with a faded blue background and deep orange text ChatGPT 4o: [creates image of a fontsheet in Avenir Next font]

User: Make a sample fontsheet with Linotype Dido ChatGPT 4o: [creates image of a fontsheet in Linotype Didot font]

Honest question: What are the legal implications of A.I.-generating an image containing a font you don’t own the license to? To my eye these are near-perfect replicas

Nice! We gotta try adding in gpt-o3-mini, gpt-4.5, and the latest gpt-4o version that's supposedly much better. Lmk if there's an easy way to contribute

Advertisement: Generate any image, any style, for free. Try Grok today. [Image of a young couple in Studio Ghibli style]

There's something really disconcerting about ripping off an artist's careful and manual style with A.I. (e.g. the viral Studio Ghibli-style image generation fad). And then it's something else entirely to advertise your money-making image generator with this unattributed slop

The original announcement of these layoffs was on the front page of the print paper on Friday, March 28.

Senior Machine Learning Engineer, A.I. Initiatives New York, NY; Remote - USA

Come work on my team! The A.I. Initiatives team at NYT blends machine learning and journalism to explore creative tools and empower responsible reporting.

We have a lot of fun and are excited to expand for the first time since I was hired last year. Remote-friendly! My DMs open for questions. #jobs

(Still not as bad as Apple telling me I haven’t finished the move ring and can still do it at like 11pm)

The Garmin watch has the audacity to give you morning messages like “Yesterday was one of the most physically demanding days in your life. Suggested workout: 90 minute hill sprint”

On the interactive on our site at least, I can promise you we’re not training an A.I. to solve puzzles based on your responses. We only store whether you get it right, use a hint or get it wrong so we can show the % of readers that got it, too.

My first deep dive story for NYT: Are You Smarter Than A.I.?

Some experts predict that A.I. will surpass human intelligence within the next few years. Play this puzzle to see how far the machines have to go. 🟦🟩🟪

Gift link: www.nytimes.com/interactive/...

Yes, that is more accurate

(Hanukkah menorahs are supposed to have eight candles on either side and one candle in the center, for a total of nine)

User: Make a glass of red wine filled entirely to the brim ChatGPT 4o: [A picture of a glass of red wine filled entirely to the brim]

User: Make me a festive scene filled with Hanukkah menorash ChatGPT 4o: [An image of multiple menorahs but only the center depicts a proper Hanukkah menorah with eight candles on either side and one in the center]

Sure, OpenAI's new vision system passes the full wine glass test but it fails miserably at the cacophony of Hanukkah menorahs test

For Trump, Biden Is an All-Purpose Target for Blame (Gift Article) A central dictum in the Trump White House is that Joseph R. Biden Jr. is to blame for just about anything and everything.

"...during the first 50 days of Mr. Trump’s presidency, he mentioned the name 'Biden' 6.32 times a day on average. It is among his most frequently used terms" (more than he said 'America')

In 2017, he mentioned 'Obama' just 35 times in his first 50 days...

— With Shawn McCreesh