Charles Salmon

old guy with a splendid wife and children

[Omar Khattab @lateinteraction]

RLMs are not sub-agents or the ability to iteratively retrieve context. I know because I trained multi-hop models for reasoning & retrieval in 2020, including compaction.*

RLMs are the simplest/purest scaffold that understands its own prompts via recursion, not via attention.

They support an extremely simple but unusual claim: Models need to be able to access their own conversations with the user and their own horizon symbolically and recursively.

The model should be only allowed to understand this long context by *writing code* that launches LLMs, and composing these into the final response.

Note that the number of LLM launches can be linear or even bigger in the context size, not a small constant number of sub-tasks. This sounds big until you remember that attention is already quadratic.

I'll have to confess that I always found (and still find) the conventional pattern of "sub-agents" rather boring. This is the superficially related structure where the model is given a special tool it can invoke by writing out prompts for and receiving the output.

Verbalizing specific individual sub-calls as tool calls token-by-token hides the internal reasoning from the main context, which is an OK outcome for sub-task delegation. But it's a completely unrelated pattern to teaching models to understand their own context/horizon recursively.

Sorry I'm a bit of a pedant for understanding concepts precisely, but this seemed needed.

*The title is quite literally "Robust Multi-Hop Reasoning at Scale via Condensed Retrieval", arXived on Jan 2nd 2021. It could work for many steps, retrieve text from a massive corpus, compact/condense its own context, and iterate further.

[Historic Vids @historyinmemes]

When a news reporter wore a grape costume to show support for a boy suspended for wearing a banana outfit to school.

In 2011, NBC4 Washington reporter Pat Collins made headlines by arriving to interview a suspended student in a full grape costume. The story began when 14-year-old Bryan Thompson, who is autistic, was suspended for wearing a banana outfit to a high school football game. During the interview, Collins, dressed head-to-toe in purple grapes, jokingly asked, “Potassium is great?”—turning a serious suspension situation into a lighthearted, human moment.

[Charles Salmon @cssalmon]

@ATabarrok the cause of, and the solution to, all of life's problems.

[sam @sam_d_1995]

republicans in montana are currently passing more progressive, sustainable, pro-housing legislation than democratic supermajorities in New York and California

[Joanne Jang @joannejang]

// i lead model behavior at openai, and wanted to share some thoughts & nuance that went into setting policy for 4o image generation.

features capital letters (!) bc i published it as a blog post:

--
This week, we launched native image generation in ChatGPT through 4o.

It was a special launch for many reasons — one of which our CEO Sam highlighted as "a new high-water mark for us in allowing creative freedom."

I wanted to unpack that a bit, as it could be easily missed by those not deep in AI or closely following our evolving thoughts on model behavior (wh… what do you mean you haven’t read the sixty-page Model Spec in your free time??).

tl;dr we’re shifting from blanket refusals in sensitive areas to a more precise approach focused on preventing real-world harm. The goal is to embrace humility: recognizing how much we don't know, and positioning ourselves to adapt as we learn.

Images are visceral
There's something uniquely powerful and visceral about images; they can deliver unmatched delight and shock. Unlike text, images transcend language barriers and evoke varied emotional responses. They can clarify complex ideas instantly.

Precisely because images carry so much impact, we felt even more heft — relative to other launches — in shaping policy and behavior.

Evolving perspectives on launching what feels like a new capability

When it comes to launching (what feels like) a new capability, our perspective has evolved across multiple launches:

1. Trusting user creativity over our own assumptions.
AI lab employees should not be the arbiters of what people should and shouldn’t be allowed to create. We’re always humbled after launch, discovering use cases we never imagined — or even ones that seem so obvious in hindsight but didn’t occur to us from our limited perspectives.

2. Seeing risks clearly, but not losing sight of everyday value to users.
It’s easy to fixate on potential harms, and broad restrictions always feel safest (and easiest!). We often catch ourselves questioning, “do we really need better meme capabilities when the same memes could be used to offend or hurt people?”. But I think that framing itself is flawed. It implies that subtle, everyday benefits must justify themselves against hypothetical worst-case scenarios, which undervalues how these small moments of delight, humor, and connection genuinely improve people’s lives.

3. Valuing unknown, unimaginable possibilities.
Maybe due to our cognitive bias against loss aversion, we rarely consider the negative impacts of inaction; some people refer to it as “invisible graveyards” although that’s a bit too morbid and extreme. There are second order or indirect impacts unlocked by a new capability: all the positive interactions, innovations, and ideas from people that never materialize simply because we feared the worst-case scenario.

How we thought about policy decisions for Day 1

Navigating these challenges is hard, but we aimed to maximize creative freedom while preventing real harm. Some examples from our launch decisions:

- Public figures: We know it can be tricky with public figures—especially when the lines blur between news, satire, and the interests of the person being depicted. We want our policies to apply fairly and equally to everyone, regardless of their “status”. But rather than be the arbiters of who is “important enough”, we decided to create an opt-out list to allow anyone who can be depicted by our models to decide for themselves.

- “Offensive” content: When it comes to “offensive” content, we pushed ourselves to reflect on whether any discomfort was stemming from our personal opinions or preferences vs. potential for real-world harm. Without clear guidelines, the model previously refused requests like "make this person’s eyes look more Asian" or "make this person heavier," unintentionally implying these attributes were inherently offensive.

- Hate symbols: We recognize symbols like swastikas carry deep and painful history. At the same time, we understand they can also appear in genuinely educational or cultural contexts. Completely banning them could erase meaningful conversations and intellectual exploration. Instead, we're iterating on technical methods to better identify and refuse harmful misuse.

- Minors: Whenever a policy decision involved younger users, we decided to play it safe: choosing stronger protections and tighter guardrails for people under 18 across research and product.

Ultimately, these considerations — coupled with our progress toward more precise technical levers — led us toward more permissive policies. We recognize this might be misinterpreted as "OpenAI lowering its safety standards,” but personally, I don’t think that does justice to the team’s extensive research, thoughtful debates, and genuine love & care for users and society.

My colleague Jason Kwon once passed onto me:
“Ships are safest in the harbor; the safest model is the one that refuses everything. But that’s not what ships or models are for.”

The future is built with imagination and adventure. As we continue our research and learn from society, we believe we can continue to find ways to responsibly increase user freedom. When (not if!) our policies evolve, updating them based on real-world feedback isn’t failure; that’s the point of iterative deployment.

Please keep sharing your feedback and creations — they genuinely help us improve!

[Charles Salmon @cssalmon]

@elonmusk Funnhy how that started under the Trump administration.

[Bryan Caplan @bryan_caplan]

If any of my friends know @elonmusk or @VivekGRamaswamy well enough to put a copy of my *Open Borders* in their hands, now's the perfect time to do so.

Yes, the secret of mass consumption is mass production. All else is a distraction.

Retweeting helps!
https://t.co/dwoX9Grltn

[Charles Salmon @cssalmon]

@McCoveyChron Polo!

[Charles Salmon @cssalmon]

@jasongay "in" Washington's Bainbridge Island?

[Charles Salmon @cssalmon]

@DouthatNYT Heretics!

[Charles Salmon @cssalmon]

@Noahpinion Orphan Master’s Son is great, but not SF.

[Charles Salmon @cssalmon]

@HanksKendyl @Popehat It’s not real. The article is 7 years old and relies on even older research. The data is “grossed up” to match alcohol sales. There are far too many assumptions to take this seriously.

[Market Urbanism @MarketUrbanism]

Combined with the end of 421-a, I think market rents in this city could get quite bad. If you aren’t in a position to buy and don’t have/can’t get a rent-stabilized lease, you should probably be thinking about what other cities you’d like to live in

[Matthew Yglesias @mattyglesias]

Making it more comfortable than ever to sit at home alone watching television? Also bad.

https://t.co/sr72YzC2bd

[Matthew Yglesias @mattyglesias]

It does seem noteworthy that the huge increase in the power of our matching tools has coincided with a *decline* in marriage rates, people having less sex, etc

[Charles Salmon @cssalmon]

@arthurbrooks I heard a Gabrielli mass there (with brass, of course). Most beautiful mass I ever heard.

[Charles Salmon @cssalmon]

@asymmetricinfo Very good piece. I’m a manager and I see each day the problems that younger workers face working remotely. You have stated the case eloquently.

[Charles Salmon @cssalmon]

@Popehat They moved to Berkeley, where they had me and the next five of my siblings. We lost her last year. I love hearing your reminiscences of LA in that time.

[Charles Salmon @cssalmon]

@Popehat Thank you for that story, sir. My mom grew up near-ish to you in La Crescenta. Her dad (after whom Humphrey Way is named) wouldn’t pay for his daughters to go to college - just Orange County CC. She saved enough to go on to UCLA, met my dad, and moved to Berkeley.

[Charles Salmon @cssalmon]

@asymmetricinfo My daughter, who's a guidance counselor at LA Unified is definitely seeing this phenomenon.