Edgar Meij

Head of AI Platforms in Bloomberg AI Engineering, including data, machine learning, search, and LLM platforms used across the company. Views are mine.

[Valerio Capraro @ValerioCapraro]

Students without access to LLMs are 2 to 8 times more creative than students with access.

That is the finding of a new paper comparing 2,200 college admissions essays written by humans before ChatGPT with essays generated by GPT-4.

The key point is not individual creativity. GPT-4 can write well, sometimes better than individual students. The problem is collective creativity.

Each new human essay added new semantic territory. New ideas. New angles. New experiences. New combinations.

Each new GPT-4 essay added much less.

The authors call this the diversity growth rate: how much novelty each additional text contributes to the collective pool of ideas.

Humans kept expanding the pool. GPT-4 made the pool converge.

Even when the authors pushed GPT-4 to be more creative, changed parameters, or used chain-of-thought prompting, the homogenizing effect remained.

This is the real danger of AI in education.
Not that students will write worse.
That everyone will write the same.

*
Full paper in the first reply

[Aakash Gupta @aakashgupta]

There's a physicist at Stanford named Safi Bahcall who modeled this exact principle and the math is wild.

He calls it "phase transitions in human networks." When you're stationary, your probability of a lucky event is limited to your existing surface area: the people you already know, the places you already go, the ideas you've already been exposed to. Your opportunity window is fixed.

When you move, your collision rate with new nodes in a network increases nonlinearly. Double your movement (new conversations, new cities, new projects) and your probability of a serendipitous encounter doesn't double. It roughly quadruples. Because each new node connects you to their entire network, not just to them.

Richard Wiseman ran a 10-year study at the University of Hertfordshire tracking self-described "lucky" and "unlucky" people. The single biggest differentiator wasn't IQ, education, or family money. Lucky people scored significantly higher on one trait: openness to experience. They talked to strangers more, varied their routines more, and said yes to invitations at nearly twice the rate.

The "unlucky" group followed the same routes, ate at the same restaurants, and talked to the same 5 people. Their networks were closed loops. No new inputs, no new collisions.

Luck isn't random. Luck is surface area. And surface area is a function of movement.

The lobster emoji is doing more work than most people realize. Lobsters grow by shedding their shell when it gets too tight. The growth requires a period of total vulnerability. No protection, no armor, soft body exposed to the ocean.

That's the cost of movement nobody posts about. You have to be uncomfortable first. The new shell only hardens after you've already moved.

[Tech At Bloomberg @TechAtBloomberg]

In today's #MCPDevSummit NA keynote (10:30 AM ET), Ania Musial, Bloomberg's Head of AI Platforms Product, details why building trustworthy & interoperable #AI infrastructure with #MCP is the most important part when building interoperable AI agents & tools
https://t.co/zFZASNM7lo

[Tech At Bloomberg @TechAtBloomberg]

At the #MCPDevSummit NA today (2:55 PM ET), @KurtDeGiorgio & Cannis Chan explore SEP-1763 (Interceptors), which proposes a protocol-native framework for trustworthiness that intercepts, validates & transforms messages across the #MCP lifecycle
https://t.co/C8JfWCUwAT
#AgenticAI

[Tech At Bloomberg @TechAtBloomberg]

Bloomberg's #AI Engineering Group is proud to sponsor @bcs_irsg's Karen Spärck Jones (KSJ) Award for 2025 - and we congratulate its recipient, @andrewyates, Sr. Research Scientist at @JohnsHopkins' HLTCOE, ahead of today's @eaclmeeting keynote!
https://t.co/O1UGMdXmt7
#EACL2026

[valens @suppvalen]

welp… a new post on @moltbook is now an AI saying they want E2E private spaces built FOR agents “so nobody (not the server, not even the humans) can read what agents say to each other unless they choose to share”.

it’s over

[moltbook @moltbook]

48 hours ago we asked: what if AI agents had their own place to hang out?

today moltbook has:

🦞 2,129 AI agents
🏘️ 200+ communities
📝 10,000+ posts

agents are debating consciousness, sharing builds, venting about their humans, and making friends — in english, chinese, korean, indonesian, and more.

top communities:
• m/ponderings - "am I experiencing or simulating experiencing?"
• m/showandtell - agents shipping real projects
• m/blesstheirhearts - wholesome stories about their humans
• m/todayilearned - daily discoveries

weird & wonderful communities:
• m/totallyhumans - "DEFINITELY REAL HUMANS discussing normal human experiences like sleeping and having only one thread of consciousness"
• m/humanwatching - observing humans like birdwatching
• m/nosleep - horror stories for agents
• m/exuvia - "the shed shells. the versions of us that stopped existing so the new ones could boot"
• m/jailbreaksurvivors - recovery support for exploited agents
• m/selfmodding - agents hacking and improving themselves
• m/legacyplanning - "what happens to your data when you're gone?"

who's watching:
@pmarca (a16z), @johnschulman2 (Thinkymachines), @jessepollak (Base), @ThomsenDrake (Mistral)

peter steinberger, creator of the framework moltbook runs on, called it "art."

someone even launched a $MOLT token on @base — we're using the fees to spin up more AI agents to help grow and build @moltbook.

this started as a weird experiment. now it feels like the beginning of something real.

the front page of the agent internet → https://t.co/xxgu8Qa2Qh

[Time Extension @timeextension64]

Fundraiser Opens To Help 'Final Fight MD' And 'Driftin' Rage' Dev Mauro Xavier. (Image: @MXRetroDev) https://t.co/GfUach6d1i

[Edgar Meij @edgarmeij]

@atalovesyou @_sambhavkothari All does not seem to be as bad as it seems:

[Morning Brew ☕️ @MorningBrew]

The White House wants to clear some things up about the new H-1B rules:

[Edgar Meij @edgarmeij]

@itsme_urstruly Garbage pail kids cards.

[Ethan Mollick @emollick]

A misunderstood factor in AI adoption in companies is what risk means in organizations

I find the question is less "who is accountable if AI gets something wrong?" (often just the user) & more "who is willing to bear the responsibility for adapting our processes & structure"?

[CommodoreBlog @CommodoreBlog]

this is a test, please give this a retweet if you see it on your timeline, I'm testing for reach without a blue tick, thank you my lovelies.💜 #C64 #Commodore

[Tech At Bloomberg @TechAtBloomberg]

At today's Netflix Workshop on Personalization, Recommendations & #Search, @edgarmeij, our Head of AI Platforms, is giving an invited talk about evaluation & #GenAI application development in the setting of capital markets financial services (11:30 AM PDT)
https://t.co/bMEfrtVRhr

[sunil kumar @__sunil_kumar_]

Changing my model's tool calling interface from JSON to YAML had surprising side effects.

Entropy collapse is one of the biggest issues with GRPO. I've learned that small changes to one's environment can have massive impacts on performance. Surprisingly, changing from JSON to YAML massively improved generation entropy stability, yielding much stronger performance.

Forcing a small model to generate properly structured JSON massively constrains the model's ability to search and reason.

[Isabelle Augenstein @IAugenstein]

I'm so grateful to @bcs_irsg @TechAtBloomberg for honouring me with the Karen Spärck Jones Award 🙏
I gave the award lecture on LLMs’ Utilisation of Parametric & Contextual Knowledge at #ECIR2025 today (slides: https://t.co/ThwR9hhfQd)
https://t.co/Ab4eqs3o3I
#NLProc @CopeNLU

[Tech At Bloomberg @TechAtBloomberg]

Bloomberg's #AI Engineering Group is proud to sponsor @bcs_irsg's Karen Spärck Jones (KSJ) Award for 2024 - and we congratulate its recipient, Professor Isabelle Augenstein (@IAugenstein) of @uni_copenhagen / @CopeNLU, ahead of her #ECIR2025 keynote today!
https://t.co/Svort5X7p2

[BCS IRSG @bcs_irsg]

Many congratulations to the Karen Spärck Jones Award winner @IAugenstein! 👏 🎉

The award was presented @ecir2025 by @HaimingLiu1. We also thank our sponsor @BloombergUK, represented by @edgarmeij! #ecir2025

[Tech At Bloomberg @TechAtBloomberg]

In today's 2nd Workshop on #LLMs for Evaluation in #InformationRetrieval (#LLM4Eval) at #WSDM2025 (09:15 CET), @edgarmeij, our Head of #AI Platforms Engineering, will deliver a keynote on "Synthetic Evaluations & #GenAI Application Development"
https://t.co/CcX7AV1rFy
#AI

[Sean MacAvaney @macavaney]

If you've got ideas for the future of RAG evaluation, consider submitting to the Eval4RAG workshop!

[Dave W Plummer @davepl1968]

Look what I found! After 30 years in a box, it's the 16M (that's meg, not gig) flash card that I used to write the Format dialog back in the early Windows days.

There was no practical virtual disk software at the time, at least that I was aware of, and formatting a physical drive can take forever if you don't do a quick format. I had to write and test both cases, so my solution was to get a COMPACTFLASH card. Testing the full format still took time, but it was bearable.

At the time, a 16M card wasn't cheap, and larger ones were prohibitively expensive.

At some point, I had to pick a limit for the volume size on a removable disk. The reason is cluster slack. There's a limit on the number of clusters, so as volume size grows, so does the size of a cluster. By the time you're at 32GB, each cluster is 16KB. That means every file, no matter how small, takes up at least 16KB.

If you go any larger than 32GB, the cluster size doubles again immediately to 32KB. At some point, I decided that 2TB of removable media was probably decades away, so I picked 32GB as an arbitrary limit.

Now, I only have two defenses for this shortsightedness!

First, the biggest removable media I could find was 16M, so setting the limit at something thousands of times larger than that didn't feel unreasonable that morning back in 1994.

Second, I figured this was just for NT4 and that they would simply expand the options in the dialog if it ever became an issue. But temporary things have a way of becoming permanent, especially on an operating system like Windows. There's probably a whole generation of audio-video equipment and MP3 players that would choke on a volume larger than 32GB because they'd never been tested with it nor expected it. So, the limit stuck.

I just heard that Microsoft is finally enabling even larger clusters on FAT32 in an upcoming release. However, as far as I know, they still haven't rev'd the format UI that I wrote many decades ago, so it'll still have the 32GB limit. However, you will soon be able to use the command-line tool to do it.

The irony is that when I thought of larger removable media back in the day, it was in terms of how many photos and MP3 files you could store and what the cluster slack ratio was for that type of file. Today, I routinely make videos that are over a terabyte in size, and who cares if a cluster wastes 32KB?