iiinova — where your digital self keeps evolving

You do realize that agentic coding is being used at scale right now? Regardless of whether you think it is a good idea or not (I personally have deep concerns regarding its effect on code quality), it is happening, and companies are benefiting from it.

100

By useful I simply mean that someone can gain benefit from using it. Right now agentic coding definitely is providing quite large benefits to some software companies (not necessarily their employees), in terms of cost to produce software. I oppose this, but acknowledge it is happening

100

Saying it is useful is different from saying it is good. Lead is useful for making water pipes, but it is not good due to poisoning. Identifying something as useful, even if it is bad, is important, so we can understand why there is demand for it

100

It's hard to say, because we don't know how much more will change in that time. If there is no limit to the software one person can get out of agentic systems, software engineering is dead anyway. If there is a limit, software demand will rise and juniors will be back to meet that demand.

Where did I say it wasn't bad? I'm saying what is happening. I even expressed dismay of it elsewhere in this thread.

100

I want to clarify that my comment on your previous thread was not meant to be argumentative in any way. However, I do understand that you receive a lot of unsolicited arguments, so if my comment came across that way, I deeply apologize.

000

Yes, the nature of business is to extract as much value as possible. Many see one bottleneck open up and think that will increase overall flow. As long as other bottlenecks exist, that isn't maintainable. Smart companies aren't expecting more flow; they are just reducing costs for current flow.

211

Yes and no. There is definitely some of that going on, and we are seeing lots of "oops" moments because of this eagerness to exploit, but I'm also seeing lower-level work being successfully automated at scale. More work for reviewers, yes, but less work overall to achieve the same results.

100

But to clarify, I definitely agree that people coming to you for validation for their moral conflicts is cringe.

000

As a very senior dev with specialized knowledge, I'm relatively safe from this disruption, but it feels unfair for me to say LLMs have no practical uses in light of the job replacement I'm already seeing for lower-level developers.

300

Kane Gregory @hobojed.bsky.social · 24/02/2026

I'm not looking for validation, just want to give data: I've happily been a software engineer for 20 years. I most definitely don't want to offload my work onto anything, let alone something so morally questionable. However, it is undeniably reaching a level that may negatively affect employment >>

140

It's a shame this sets the cost to 1 rather than reduces it by 1, because as it is, it literally makes DH's hp strictly inferior to Druid's

100

Kane Gregory @hobojed.bsky.social · 23/02/2026

While it may not be what leads to true AGI, my worry is that what it does lead to is good enough to displace most humans economically in a very short span of time. I'm not sure how civil society survives that, and by extension, how geopolitical relations don't completely break down, either.

100

Kane Gregory @hobojed.bsky.social · 12/02/2026

It was likening LLMs to cold readers that finally let me truly understand @emilymbender.bsky.social's observation that the only meaning in an LLM's output is provided by the user, and not the LLM itself.

041

Kane Gregory @hobojed.bsky.social · 30/01/2026

When I used to follow competitive Magic (a long time ago, now), Kai stood out from the crowd. Not only because of his amazing achievements, but also because it was clear he genuinely was a really great guy. There was so much negativity in the scene, and he rose above it

000

Kane Gregory @hobojed.bsky.social · 25/11/2025

Sadly, he takes the wrong lesson from this. Instead of pondering, "perhaps they aren't working the way I think they are working," he instead goes straight to "I guess the training strategies we use make them too single-minded."

000

Kane Gregory @hobojed.bsky.social · 13/11/2025

If they don't know about a subject and are unwilling to do research, then perhaps they shouldn't be writing about it? They then had the audacity to actually ask you to do their work for them!

Kane Gregory @hobojed.bsky.social · 27/10/2025

But then, so does "Zee's dead, baby. Zee's dead."

3210

Kane Gregory @hobojed.bsky.social · 24/10/2025

I think a big problem with it is that Creature of Madness just does the job so much better. Reliably dropping a Dark Gift minion on turn 3 is so much better than possibly dropping one on 4

000

Kane Gregory @hobojed.bsky.social · 09/10/2025

Personally, I believe that procreation is not, in and of itself, a moral concern. Although, of course, like with any action, there are secondary concerns to consider. Many pro- and anti-natalists focus their arguments on these secondary concerns, which is fine. Replacement anti-natalists do not.

030

Kane Gregory @hobojed.bsky.social · 03/10/2025

Exactly! The Turing Test assumes humans are good at differentiating text with meaning from text without meaning. Our hubris told us we were, but we should have known better. The fact that humans are often fooled by cold readings and other cons should have been warning enough that we were wrong.

000

Kane Gregory @hobojed.bsky.social · 24/09/2025

That's why I wait for tech to be proven. It either: 1) Gives cheap and fast productivity boosts: I can integrate and catch up quickly 2) Gives expensive or slow productivity boosts: Nobody got significantly ahead while I waited 3) Gives no productivity boost: I didn't waste time/money on it

011

Kane Gregory @hobojed.bsky.social · 24/09/2025

Algorithmic social media streams do not work in favor of the user. Many people try to solve that with block lists, but that puts the power over your attention into someone else's hands. I prefer to just avoid algorithmic social media altogether, and use follow-only streams where possible.

040

Kane Gregory @hobojed.bsky.social · 17/09/2025

I haven't played the game in weeks now, and this balance change is not bringing me back. The game has been too stale for too long. It doesn't matter if it is balanced, if it's just the same stuff we were playing half a year ago.

000

Kane Gregory @hobojed.bsky.social · 11/09/2025

I have a feeling that in 10 years, I may need to launch the Center for the Alignment of Alignment Centers for AI Alignment Centers.

1141

Kane Gregory @hobojed.bsky.social · 09/09/2025

The most popular LLMs all produce sycophantic output (which is a big part of why people are becoming addicted to them), so a simple prompt like "Write a glowing review for this book" would likely be enough to pattern-match to blurbs like these.

000

Kane Gregory @hobojed.bsky.social · 03/09/2025

I find it hard to believe they didn't know this mini set would be very low-impact, so the question is, why were they fine with that?

000

Kane Gregory @hobojed.bsky.social · 26/08/2025

It's easy to block certain words, but any content moderator knows it is hard to automate blocking the discussion of specific topics. The words you block can have legitimate uses ("I'm hosting a murder mystery party...") and the topics you don't want discussed can use words you didn't think to block.

100

Reposted by Kane Gregory

Kashmir Hill @kashhill.bsky.social · 26/08/2025

Adam Raine, 16, died from suicide in April after months on ChatGPT discussing plans to end his life. His parents have filed the first known case against OpenAI for wrongful death. Overwhelming at times to work on this story, but here it is. My latest on AI chatbots: www.nytimes.com/2025/08/26/t...

nytimes.com

A Teen Was Suicidal. ChatGPT Was the Friend He Confided In.

10845961702

Kane Gregory @hobojed.bsky.social · 26/08/2025

This isn't just about benevolence. They couldn't make it reliably safe if they tried, because of how LLMs work. The benevolent thing to do would be to not market them as chatbots at all, and not to train them to respond with anthropomorphic language.

2160

Kane Gregory @hobojed.bsky.social · 22/07/2025

I would argue that they need to be carefully boxed in even if they are only attached to humans. Even simple chatbots are causing a lot of real-world problems due to the human mind being predisposed to believe their words have meaning

020

Kane Gregory @hobojed.bsky.social · 15/07/2025

Buff-wise, I'm not so sure. Yes, the quests aren't doing well, but the meta usually sucks when quests are doing well. Some could have minor adjustments to pull them up to a playable level (many are hot garbage as it stands), but they really don't need to be tier 1.

000

Kane Gregory @hobojed.bsky.social · 15/07/2025

More than anything, I think Menagerie Jug needs a change. After that, Careless Crafter and/or Resuscitate should probably change. Otherwise, OTK Priest will dominate. Then the obvious, but not as important, Murloc Paladin and Loh. Both need a slight tone down, but probably nothing too serious.

110

Kane Gregory @hobojed.bsky.social · 15/07/2025

I really enjoyed it. Mr. Terrific in particular was a pleasant surprise. Hopefully, there are future projects where he gets more time to shine

000

Kane Gregory @hobojed.bsky.social · 30/06/2025

I remember learning about this competition when I was at University in 2001. It was a well-established event even then!

000

Even in the case of pure memorization, LLMs do not learn the same way. If you show me a block of random words, I may exactly memorize a small number of them very fast, but I would not be able to reproduce the word frequencies in a large piece of text. The reverse is true for an LLM.

200

I want to make that comparison because the court case, which this thread is about, was about that. The court case isn't about what LLMs may be able to do. It is specifically about how LLMs learn now.

000

How are any of these questions relevant to the original point, which is specifically about how humans and current LLMs respectively learn from text. I've never stated that the underlying mechanisms are different, just that the process is different (and that difference is important)

000

How many times are you going to change the topic of discussion. This is not about LLMs potential, this is about how they learn right now.

Once again, this is not what we are talking about. The claim was that current LLMs learn the same way humans do. Whether they could be updated to do so in future is irrelevant to the fact that they currently do not do so.

000

If they could do it, they would have. The reason they haven't is because they are hard problems to solve. But none of those things relate to the topic here, which is how current LLMs do not learn the same way humans do.

000

None of this is relevant to the original topic: Current LLMs learn by us directly updating their weights to make the output closer to what is expected; humans learn by contemplating meaning and then updating their understanding (which may be neural weights). That is what the case is about.

000

Sure, that is arguably true (although we don't know the full extent of human learning mechanisms), but my point is that humans reflect on the meaning of a text before updating. This is why we can permanently learn a whole new concept from a single sentence, and LLMs cannot.