Basically, if you tell a model "You're an absolute moron, of course that's wrong!", will it give better or worse results? How much of that response will it absorb into its persona (like some humans tend to do)? Will it try to give "safer" responses to avoid negative feedback? How much of the associated behavior can be attributed to RLHF (e.g. like the sycophantic nature of LLMs)? How much can be attributed to training data?

Obviously this will vary by model and training, but I'm trying to get a general understanding.

I recall seeing related outcomes in some of Anthropic's studies, but I'm not sure how much of this particular aspect was studied.

Comments URL: https://news.ycombinator.com/item?id=48275667

Points: 1

# Comments: 0

"Here's AI. Figure out how we can make money from it. We're adding it to your performance reviews"

Basically, here's a solution. Find problems for it.

This is also why one of my instructions to coding agents is that they adhere to established coding and testing patterns, even where they appear to be sub-optimal.

https://news.ycombinator.com/item?id=47903126

I wondered how far I could get by just choosing a random open problem and throwing it at LLMs.

Disclosure: I have no idea what the problem even refers to, let alone whether or not the output is even remotely correct. My interest is purely for testing capabilities of various models, curiousity, and entertainment.

Problem: https://www.erdosproblems.com/691

  Given A\subseteq \mathbb{N} let M_A=\{ n \geq 1 : a\mid n\textrm{ for some }a\in A\} be the set of multiples of A. 
  Find a necessary and sufficient condition on A for M_A to have density 1.

Once it generated a proof, I asked Opus 4.7 to review it, and then entered the review into DeepSeek which made edits, corrections and refinements. This back-and-forth continued till Opus 4.7 was reasonably happy. At that point, I called in Gemini 3.1 Pro Preview, which raised issues which Opus missed. Opus acknowledged the feedback, and then I placed its feedback into Deepseek for a final round. Essentially, what Opus says Deepseek generated was a "clean exposition of a D[avenport]-E[rdos] corollary", not a new result. In all likelihood this result may already be known (Deepseek was not allowed to use the internet for this phase), or even wrong.

In "simple" terms:

  The argument actually proves a stronger fact for every set \( A \) of natural numbers:  
  The upper density of the set \( M_A \) equals the largest possible lower density you can get from finite subsets of \( A \), and that also equals the lower density of \( M_A \).
  When the upper density is 1, it forces the lower density to also be 1, so the natural (ordinary) density exists and equals 1 automatically, without needing any extra conditions.
  The only non-basic part of the proof is the Davenport–Erdős theorem; everything else is simple.

- These new models seem to be surprisingly capable especially when used to in conjunction with each other, even with fairly simple prompts

- I am quite impressed by Deepseek. I'm going to review its coding ability, and may even switch completely from Anthropic

- This was a genuinely interesting exercise, even if I have no idea if any of it is correct or useful

Some other observations:

- Opus was really fast at reviewing Deepseek's output. Literally seconds

- Gemini had trouble figuring out what "Erdos 691" referred to

- The free version of ChatGPT of generated mostly useless output. I didn't include it.

Chat links below:

https://chat.deepseek.com/share/hpguvrhcxn226bi3hn

https://claude.ai/share/4f3ccad1-d862-4e37-8333-8a1ebd84b38f

https://aistudio.google.com/app/prompts?state=%7B%22ids%22:%...

Comments URL: https://news.ycombinator.com/item?id=47917671

Points: 1

# Comments: 0

What is the current pattern in other industries?

Does the pattern exist elsewhere in the world?

Of course. But you shouldn't run retros that are focused on it.

e.g.

- 45 minute "standups" (!?)

- PI "planning" that consisting of deadlines and glorified multiplayer MS Paint

- Rigid adherence to ceremonies or processes that add zero value

- Retros that focus on complaints and venting with no actionable outcomes

- etc etc

Every time I've introduced Agile to a team or project that was new to it I was always met with skepticism. But 6 months down the line noone on the team/project wanted to go back to the "old" way of working. I don't even really care about any text book definitions. These are the only things we try to stick to:

- Short, daily standups

- Planning based on risk reduction

- Estimates based on complexity (ties in with risk reduction)

- Actionable retro items

- User demos every sprint (makes it easier to pivot - users rarely know what they want)

https://www.youtube.com/watch?v=huAIfzUoyhU&t=121s

https://www.youtube.com/watch?v=fLzEX1TPBFM

Just because a country generates 100% of its energy from renewables, it doesn't mean that its enough to power the entire or even majority of the country. Case in point: DRC. I believe only half of the population has access to electricity. It's been a while since I've looked into continental stats, but a quick Google search suggests the situation hasn't changed that much in the last few years.

I may be out of the loop, but how was Bitwarden not "good" to users? Does this relate to the recent price increase?

I've used Windows since 3.1. Win 11 was the straw that broke the camel's back. I moved to CachyOS a few months ago and I honestly can't find a reason to switch back.

Yes. I mean... of course he was?. Firstly, I had already gone through this process with multiple LLMs, from various perspectives, including using Deep Research models to find out if any other businesses faced similar issues, and/or if products existed that could help with this. That lead me down a rabbit hole of data science products related to regulatory reporting of a completely different nature which was effectively useless. tl;dr: Virtually all LLMs - after understanding the context - recommended us doing thing we had already been urging the business to do - hire a Technical BA with experience in this field. And yes, that's what we ended up doing.

Now, give you some ideas about why his idea was obviously absurd:

- He had never seen the SP

- He didn't understand anything about regulatory reporting

- He didn't understand anything about financial derivatives

- He didn't understand the difference between Transact SQL and ANSI SQL

- No consideration given to IP

- etc etc

Those are the basics. Let's jump a little bit into the detail. Here's a rough snippet of what the SP looks like:

   SELECT
    CASE
    WHEN t.FLD4_TXT IN ('CCS', 'CAC', 'DEBT', ..... 'ZBBR') THEN '37772BCA2221'
    WHEN t.FLD4_TXT IN ('STCB') AND ISNULL(s.FLD5_TXT, s.FLD1_TXT) = 'X' THEN 'EUMKRT090011'
    END as [Id When CounterParty Has No Valid LEI in Region]
   -- remember, this is around 5000 lines long ....

So... you're not just dealing with incredibly unwieldy and non-standard SQL (omitted), noone really understands the business rules either.

So again... yes he was entirely wrong. There is nothing "trivial" about refactoring things that noone understands.

a) They asked an LLM

b) "This is what all our competitors are doing"

c) They saw a video on Youtube by some big influencer

d) [...insert any other absurd reason...]

True story:

In one of our recent Enterprise Architecture meetings, I was lamenting the lack of a plan to deal with our massive tech debt, and used an example of a 5000 line regulatory reporting stored procedure written 10 years ago that noone understood. I was told my complaint was irrelevant because I could just dump it into ChatGPT and it would explain it to me. These are words uttered by a so-called Senior Developer, in an Enterprise Architecture meeting.

Who are these people? What is the analog for this corner of the market? Context: I'm a 47y/o developer who has seen and done most of the common and not-so-common things in software development.

This segment reminds me of the hoards of npm evangelists back in the day who lauded the idea that you could download packages to add two numbers, or to capitalise the letter `m` (the disdain is intentional).

Am I being too harsh though? What opportunity am I missing out on? Besides the potential for engagement farming...

EDIT: I got about a minute into Fireship's video* about this and after seeing that Whatsapp sidebar popup it struck me... this thing can be a boon for scammers. Remote control, automated responses based on sentiment, targeted and personalised messaging. Not that none of this isn't possible already, but having it packaged like this makes it even easier to customise and redistribute on various blackmarkets etc.

EDIT 2: Seems like many other use-cases are available for viewing in https://www.moltbook.com/m/introductions. Many of these are probably LARPs, but if not, I wonder how many people are comfortable with AI agents posting personal details about "their humans" on the net. This post is comedy gold though: https://www.moltbook.com/post/cbd6474f-8478-4894-95f1-7b104a...

[*] https://www.youtube.com/watch?v=ssYt09bCgUY

It's unfortunate that so many people end up parroting fanciful ideas without fully appreciating the different contexts around software development.

Heh. I've been saying that since I was on Mandrake in the early 2000s. This is just what the Linux landscape is like.

That said, I'm generally not easily impressed, especially by random *nix distro 347, but CachyOS is surprisingly good. I've finally switched full time from Windows. I don't even need VS anymore because Rider is x-platform.