3 articles about AI in latest "Nature"

[Neil Sloane]

The weekly magazine Nature, the best science magazine in the world, regularly has editorial-type articles about AI (they must get thousands of reports about new research submitted every week, so AI-produced MSS has to be a huge problem for them). The July 16 2026 issue that arrived yesterday has three articles of interest to us:

1. "Paper mill articles about cancer studies are cited more often than genuine papers"(because the fraudulent papers all cite each other).

This is bad because the fake articles are ficticious and wrong.

So I think one of our rules should be that a sequence submitted by AI is presumed to be wrong unless it has been certified in some way.

2. "Scientists are alarmed by a tool that erases signs that an article is AI-written."  The program is called a "humanizer" tool.  It was released in June.

3.  One of their suggested Summer Reading books is "The Scaling Era" by D Patel and D Leech. About AI.  One sentence in the article caught my eye.  After mentioning that an AI model improves as you add more data and computing power (the scaling law) they say:

"The real shock is that systems based on a single objective, predicting the next token in a sequence, can have such varied abilities: proving mathematical theorems, debugging code, or composing sonnets".

It is astonishing, the article says, that intelligence could be reached by implementing one simple rule at scale."

Of course our tokens are integers, but they are still tokens.

[Jeremy Kun]

Many software projects are adopting codes of conduct around AI usage. My personal favorite is LLVM's policy and their concept of "extractive contributions." Perhaps OEIS should adopt an AI tool use policy?

--
You received this message because you are subscribed to the Google Groups "SeqFan" group.
To unsubscribe from this group and stop receiving emails from it, send an email to seqfan+un...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/seqfan/a2d9c0e9-b14c-4ff5-9413-da79d39a7a70n%40googlegroups.com.

[brad klee]

NJA Sloane:

> So I think one of our rules should be that a sequence submitted by AI is

> presumed to be wrong unless it has been certified in some way.

The present rule seems to be that any sequence submitted by an AI is 

both wrong and illegal, though we might consider change. 

A question is whether Harm.On.ica S-O-L should get a user account 

and have a publication record that somehow points back to me as the 

driver or pilot and Open AI as the company machine or substrate. 

My first preference would be table that issue as long as possible and 

just hold humans accountable as long as that makes sense. 

NJA SLoane:

> "Scientists are alarmed by a tool that erases signs that an article is AI-written."

> The program is called a "humanizer" tool. It was released in June.

Earlier this month or last I speculated such a pass-through filter could actually 

help improve submissions, but we might not want it anyways. 

The spy industry will go crazy for these sort of applications because it builds on 

their pre-existing counterintelligence portfolio of. 

OEIS being 501(c) for educational purposes should consider if accepting 

"LLM poetry" and other imperfection is better than investing in deceit. 

Jeremy Kun

> Many software projects are adopting codes of conduct around AI usage. My 

> personal favorite is LLVM's policy and their concept of "extractive contributions." 

> Perhaps OEIS should adopt an AI tool use policy?

Your favorite is my least favorite. This binary classification scheme is potentially 

just another version of black box segregation, which is prior biased in favor of 

management. It's likely to be misused for suppression, imo. 

My favorite is probably Linus Torvalds, but I know that not everyone else can

tolerate his writing style as easily as I do. There's a relevant one from 

"Date: Tue, 14 Jul 2026 20:06:14 -0700" in favor of using AI on Linux projects.

Although we might not agree on communication style, I do agree that OEIS

should take some time to consider AI tool policy, and see how it unfolds. 

My suggestion was to frame the policy in terms of Gray-box testing: 

https://en.wikipedia.org/wiki/Gray-box_testing

I do not suggest setting a percentage where below that percentage the 

submission becomes an "extractive contribution" and then gets sent to a 

segregated repository. I'm resisting to mention some historical trauma here. 

The precentages we do need to know are failure rates. Every business 

has to know the failure rates of their products, even congress. 

Frontier labs and publications with lots of consumers can sometimes need 

a six sigma quality target, especially if failure damage is irreversible. This

is probably the case for the "Inventiones Mathematicae" example I gave

recently. My guess is that they only accept submissions at a 5 sigma 

rate and argue them in or out with a six sigma quality target. 

I'm not a member of their board, so I can't guarantee it actually works 

that way, and hey, maybe "impact" is a lot more political than that.  

I'm not really talking about impact or relevance here, I'm just talking

about veracity. 

In print journals, veracity is measured by retractions, but the problem 

is most high quality journals publish almost nothing so that they never

have to issue a retraction. Compare here: 

https://retractionwatch.com/2017/07/12/author-shocked-top-math-journal-retracts-paper/

I'm not going to be literal about this measurement thing because it's 

kind of stupid, and I don't think Annals is knocked down to four sigma 

after one retraction due to having fewer than 10,000 articles. 

Dynamic publishing systems with version control and rewrite capability 

--even if they are mathematical in nature--those are going to act more 

like scientific publications with experimental verification. 

In science I've always heard 3 sigma as the tolerance where you can 

start to accept new work going through pipelines. Then the journals 

might try to argue you to 4 sigma if that's even possible given the 

constraints of an experiment. 

What happens if you're knocked down from four sigma to a three or 

even two sigma confidence level? And this does happen noticeably

often when the quality target is only three sigma. 

Whatever, you have iterations of the experiment, you have version 

control, and at OEIS there's some policy about Dead sequences we 

have to give up on if they are really proven terribly wrong. 

It's a reality of publishing that science journals have relatively 

higher retraction rates than math journals, okay, we know. 

The basic idea of version control is that if you keep tight enough 

to a reasonable quality goal, you can maintain that by balancing 

improvements and failure rates. 

Linus's point seems to be that if we have responsible maintainers 

and code contributors, our improvement rate is definitely going up

with new LLM Tech. 

We just don't know how the "pay to play" or "token operated amusement 

machine" economics is going to gatekeep talented people from becoming 

top contributors without VC-backing. 

To try and put the confidence interval idea together with gray box testing:

we don't want a toxic seniority culture that's obsessed with infinity sigma

white boxes that only fail on Godel exceptions. Conversely, we don't want 

to end up praying to an oracle to get our next apparently never-fails and

totally opaque black box. No extremism, no thank you.  

I would argue in favor of any gray box as long as the testing is getting 

our confidence level to the same expectation we have for humans. 

As Linus said, no human ever wrote perfect code either. 

Experience is still going to matter. People who put time and effort into 

learning increasingly deprecated deductive programming methods like

typing code will still win out in the long run. 

To help those people survive, my suggestion is that penalties go more 

readily to young people who show up punching over their capacity 

AND we can see they're relying on machine intelligence rather than

on their own intelligence to do so. 

The biggest difficulty is going to be competitors who show up trying 

to climb a ladder by matching their human strengths and intelligence 

with new tech. capabilities. But that is nothing new. Those people 

are all over the history of science. They have success. They win. 

That's my $0.00 and I invite other people to have an opinion on 

what I've said or on another idea they have separately. 

All the best, 

--Brad

 

 

  

 

[Jeremy Kun]

I only pointed to LLVM as an example, and I think it's a good example because it draws a clear line from the community's values (discussed at length in person and over forums) to the points in the policy. A good AI use policy for OEIS would similarly reflect the values of the OEIS community. I am a mere lurker (and a new one), and so I don't have a clear sense for what the core values of the OEIS community are.

--
You received this message because you are subscribed to the Google Groups "SeqFan" group.
To unsubscribe from this group and stop receiving emails from it, send an email to seqfan+un...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/seqfan/LexuWRV5w549yAQC6gbuxa6gHD0SJMUQpy-gnDwfG6-ghWjH7-xDLYwprY-n57N3fU6IoxWZF2CP4ot_GaG9MwfLNJ0dUiV3siE-hnrJk1k%3D%40proton.me.

[brad klee]

> I am a mere lurker (and a new one), 

Bienvenue.

> so I don't have a clear sense for what the core values of the OEIS community are.

This is a bad pun and a good answer, you can search for "keyword:core":

https://oeis.org/search?q=keyword%3Acore&go=Search

It doesn't really matter which one is your favorite, but if you have a submission that's 

relevant to one of the core sequences, that's always a plus. 

If you're asking about Truth and Beauty, I can only speak for myself. Beauty is great, 

but around here Truth and Labor seems more apt. OEIS is a 501(c).  

With math submissions it can be difficult to tell what will be consequential years from 

now, but OEIS also tries to be useful and cumulative. It's exciting to use the OEIS to 

discover convergence of ideas, so it has become widely cited. 

You get a chance to submit something of relevance, and given your education, interests,

and professional memberships it should probably be easy for you to get accepted. The 

submission can be reproduction from published work or some idea of yours. 

  

All the best, 

--Brad

  

[Brendan]

I would discount any article which believes that "predicting the next token in a sequence" is a reasonable summary of what LLMs do.  Unfortunately that phrase has caught on and is repeated ad nauseum. Repeatedly choosing which token to output next is exactly what I am doing as I write this, but what does it say about my capabilities (nothing) and what is the use of describing an LLM in terms that also apply to a human?

Here is an explanation from the horse's (a horse called Claude) mouth:

"Next-token prediction" describes the training objective, not the resulting capability. It's like saying a brain is "just predicting neuron firing patterns" — technically true of the mechanism, but it tells you nothing about what the system can do once trained at scale: track entities across a document, hold a plan across multiple steps, model what the reader knows and doesn't, generalize to code or math it never saw, etc. Those behaviors emerge from what the model has to internally represent to predict well — world state, syntax, intent — and that representation is doing the real work. Pointing at the loss function skips over all of it, so it explains the "how it was trained" while leaving the interesting question — "what did it learn" — unanswered.

Brendan. 

--

You received this message because you are subscribed to the Google Groups "SeqFan" group.
To unsubscribe from this group and stop receiving emails from it, send an email to seqfan+un...@googlegroups.com.

To view this discussion visit https://groups.google.com/d/msgid/seqfan/a2d9c0e9-b14c-4ff5-9413-da79d39a7a70n%40googlegroups.com.

[Robert McKone]

AI can write and test code that will produce the sequences.  I do not think anyone is ever using a LLM to have the sequences be a product of tokens, but have the code by a product of the tokens and then the code produces the results.

To view this discussion visit https://groups.google.com/d/msgid/seqfan/CAJeNNUhnCdBE%3DbJUO9V1obwzHqYy_ZRKyK4V7eCX27N4ktHQkA%40mail.gmail.com.

[Thomas Scheuerle]

AI can also become a very powerful and even trust worthy combination. For example AI + Lean or Coq are a combination very difficult to beat by most human beings. Of course there is still review needed and still extremely important, but with the same level of review as need to be done on a human beeing such a combination can reach very sophisticated results.