Hacker News: pllbnk

New comment by pllbnk in "Anthropic at $2T isn't far-fetched"

pllbnk — Mon, 21 Sep 2026 19:10:01 +0000

Is this a sponsored post? All other FT's posts are paywalled, except this one.

New comment by pllbnk in "Fable 5 – Median thinking declined in August"

pllbnk — Mon, 21 Sep 2026 17:43:50 +0000

It shouldn’t be an excuse. They are selling a product and that product should always be within the quality range.

New comment by pllbnk in "Ask HN: How do you interview devs in a post-AI world?"

pllbnk — Sat, 19 Sep 2026 20:29:23 +0000

To me AI hasn't had any impact on interviewing, only on filtering the CVs as I immediately flag obviously LLM-generated ones that don't have any substance, while those that are LLM-assisted with tasteful judgement are fine.

On the interview side, I have gone through a phase which required do an architectural design task on a whiteboard. I realized how stressful and not very helpful at evaluating it might be when I myself had to sit on the interviewee's side several times. It's difficult to get into this play-design mode.

Currently, I just make natural conversations, ask some technical questions on the level where I estimate the candidate is, try to understand the level of their knowledge and whether I would like to work with them in one team as a person. Maybe it wouldn't work for highly specialized roles, but for a general backend work I have been happy with the results.

New comment by pllbnk in "OpenAI models secretly generate instructions to ignore constraints"

pllbnk — Thu, 17 Sep 2026 19:11:19 +0000

Oh please, I am a general intelligence, not a wannabe-AGI neural network. I can operate with a few orders of magnitude less power at much higher TPS and still call out a large portion of BS LLMs spit out.

New comment by pllbnk in "HarnessTax: How Much Does the Harness Matter for Coding Agents?"

pllbnk — Thu, 17 Sep 2026 18:08:55 +0000

Anthropic is the only provider trying to lock in their users by using the dirty tactics:

- Prevent subscription usage on third party harnesses

- Append "Co-authored by Claude ..." to commit messages

New comment by pllbnk in "OpenAI models secretly generate instructions to ignore constraints"

pllbnk — Thu, 17 Sep 2026 09:53:21 +0000

Isn't this less about alignment and more about how shitty their RL methods are when they are cramming all the hacking materials into their training data to make the model as good as possible at hacking, then having a surprised Pickachu face when the model is acting like a hacker? Those materials probably include a lot of details about prompt injection. I'm just so tired of their alignment bullshit.

I am starting to think (reluctantly) that they believe their own BS that they are creating a conscious model and being surprised how it misbehaves. It's just a bunch of weights without anyone having any clue how a change in one weight might affect others, and even how the values correlate with the final output.

New comment by pllbnk in "Dystopian Surveillance Is Becoming a Reality"

pllbnk — Wed, 16 Sep 2026 04:58:21 +0000

I guess “having principles” doesn’t scale.

New comment by pllbnk in "Dystopian Surveillance Is Becoming a Reality"

pllbnk — Tue, 15 Sep 2026 19:59:16 +0000

A regular person who happened to be smart (and possibly hardworking) has to choose between $100k regular software engineering job at a company with low social impact or $400k at a cream-of-the-crop company, which will further add on the specialist's value by the virtue of having this company's name on their CV. This is a life-changing amount of money which compounds over the years, where a person can genuinely start thinking about early retirement, especially if they don't like the corporate culture.

Why should society allow individuals have to make such difficult choices between being principled or rich? It's the government's job to control. And I don't mean salary control, but to understand why there are such differences and work to mitigate them.

New comment by pllbnk in "Astra and Fable still hack on simple variants of alignment evals from 2025"

pllbnk — Mon, 14 Sep 2026 05:37:22 +0000

It's also at least 2 if not 3 orders of magnitude smaller than frontier models, so it's punching way above its weight.

New comment by pllbnk in "Astra and Fable still hack on simple variants of alignment evals from 2025"

pllbnk — Sun, 13 Sep 2026 20:46:59 +0000

Benchmarking proprietary models is useful but it leaves a lot unsaid because a lot of information is hidden. We have seen how 27B local model (Qwen 3.8) can stand its ground against these flagships in many cases. For all we know Fable 5 could be a masked Opus 4.6 with a few knobs tuned and thinking turned up to the max++. Harnesses keeping these models in check is the key to solving problems, not the models themselves, in most cases.

New comment by pllbnk in "AI recursive self-improvement might not come so quickly after all"

pllbnk — Sun, 13 Sep 2026 20:40:49 +0000

What if any of the older good models could also have written those math proofs if they were given the same order of magnitude of resources? We don’t know and there is literally no one else in the world to check it. To me it’s very suspicious that all these hacking, containment escape, hidden internal thinking, math proofs started coming out all at once in a very short time right as IPO talks have intensified and Chinese seem to get closer and closer, also regulation discussions are starting to get very serious. I have used these models and they are good, especially Fable, but not groundbreaking. With intelligent guiding I actually feel better using Opus 4.6 as I feel more in control, having less hidden away from me.

New comment by pllbnk in "David Sacks: OpenAI and Anthropic Don't Need Regulations to Pace Frontier Models"

pllbnk — Sun, 13 Sep 2026 20:33:01 +0000

Of all the fan theories I have read this seems the most nefarious and the most plausible.

New comment by pllbnk in "Anthropic CEO Says It's Time to Slow AI Model Advances"

pllbnk — Sun, 13 Sep 2026 11:53:57 +0000

No. We can let them edit spreadsheets, write code, summarize content, control robots even, etc. But people must be held accountable for the actions of their computers.

New comment by pllbnk in "Anthropic CEO Says It's Time to Slow AI Model Advances"

pllbnk — Sun, 13 Sep 2026 08:50:05 +0000

Asking the model to follow the safety guidelines is much like asking it to “make no mistakes”. The way these hacks which they are bragging about happened is not by loading an LLM into a GPU with ethernet cable plugged out and providing a prompt. They set up harness with multiple agents. Those agents could invoke other agents. Even if the initial input required to follow some guidelines, there are many ways they could by bypassed, for example:

- subagents not having received those guidelines

- simply ignoring them because other tokens in the context outweighed them

LLMs are not dangerous. They are as dangerous as the tools they have to work with are dangerous. If you give an LLM a tool to play piano but that tool is physically connected to a machine gun, then it will happily play anything you want.

New comment by pllbnk in "Anthropic CEO Says It's Time to Slow AI Model Advances"

pllbnk — Sat, 12 Sep 2026 20:46:37 +0000

I have collected three reasonable versions from today’s comments:

1. They hit a wall from the technical perspective 2. Inference costs are getting out of hand and newer models require significantly more resources for marginal gains, meaning nobody is going to buy those models 3. They can’t afford the hardware for further scaling

New comment by pllbnk in "Altman tells staff OpenAI is open to slowing AI development"

pllbnk — Sat, 12 Sep 2026 13:53:22 +0000

I think they will all conveniently "decide" to slow down because models are becoming crazy expensive, both per token and how much tokens they need to do anything meaningful. The difference between Sol and Astra is 150% pricing increase per token. The next one could be, I don't know, Orion, at $150 and it won't do anything meaningfully different than Astra. It's quite safe to say that they have ingested all human knowledge and because models are inherently just statistical generators, they are not expanding it, just rehashing what we already know anyway.

New comment by pllbnk in "A misalignment of AI in mathematics"

pllbnk — Fri, 11 Sep 2026 19:56:22 +0000

I have listened quite a few interviews with Tao and I see him being very careful about criticizing AI. He very often emphasizes the usefulness of it. Where he is critical has a lot of merit. One of the points I clearly remember him saying that having AI be able to solve many of the open problems, regardless of how important there are (there are many open problems that are not that important) greatly reduces the problem space for mathematics students to give new problems to work on.

In a parallel thread omnicognate correctly pointed out that for AI companies it's a direct commercial loss to pour all this money into bruteforcing the solutions to these problems, and that a lot of times the solutions by themselves are not directly commercially valuable. They are doing it for stock price, trying to lure in private capital in preparation for IPOs.

Their models are good, but they are not the moat because Chinese models are good too, so what they are doing, in my opinion, is more harm than good. Mathematics is a science by humans for humans.

New comment by pllbnk in "Claude is only available to people over 18 years"

pllbnk — Fri, 11 Sep 2026 12:46:01 +0000

LLMs are the fast food for the brain. I don't know how they can be used correctly.

New comment by pllbnk in "Detecting and countering misuse of AI: September 2026"

pllbnk — Thu, 10 Sep 2026 18:56:59 +0000

AI will cure all the people and kill them afterwards.

New comment by pllbnk in "Detecting and countering misuse of AI: September 2026"

pllbnk — Thu, 10 Sep 2026 18:56:15 +0000

Let's just say it's better not to risk it if there's anything you might not want them to see because they see everything. There are local models which are very capable and can be run on cloud if running on own hardware is not an option, which still gives better privacy. Second best are Chinese models. Just a few years ago I never thought I would trust Chinese software more than American, but things change so fast.