Hacker News: __alexs

New comment by __alexs in "This year’s insane timeline of hacks"

__alexs — Mon, 13 Apr 2026 16:57:03 +0000

Anthropic's marketing team are terrifyingly good. I wonder if Opus came up with this plan?

New comment by __alexs in "I still prefer MCP over skills"

__alexs — Fri, 10 Apr 2026 13:48:49 +0000

It cannot do "anything" with the tools. Tools are very constrained in that the agent must insert into it's context the tool call, and it can only receive the response of the tool directly back into its context.

Tools themselves also cannot be composed in any SOTA models. Composition is not a feature the tool schema supports and they are not trained on it.

Models obviously understand the general concept of function composition, but we don't currently provide the environments in which this is actually possible out side of highly generic tools like Bash or sandboxed execution environments like https://agenttoolprotocol.com/

New comment by __alexs in "I still prefer MCP over skills"

__alexs — Fri, 10 Apr 2026 12:17:12 +0000

The agent cannot compose MCPs.

What it can do is call multiple MCPs, dumping tons of crap into the context and then separately run some analysis on that data.

Composable MCPs would require some sort of external sandbox in which the agent can write small bits of code to transform and filter the results from one MCP to the next.

New comment by __alexs in "Claude mixes up who said what and that's not OK"

__alexs — Thu, 09 Apr 2026 11:50:50 +0000

This hasn't been the full story for years now. All SOTA models are strongly post-trained with reinforcement learning to improve performance on specific problems and interaction patterns.

The vast majority of this training data is generated synthetically.

New comment by __alexs in "Claude mixes up who said what and that's not OK"

__alexs — Thu, 09 Apr 2026 11:28:31 +0000

I think OpenAI and Anthropic probably have a lot of that lying around by now.

New comment by __alexs in "Claude mixes up who said what and that's not OK"

__alexs — Thu, 09 Apr 2026 10:23:42 +0000

The models are already massively over trained. Perhaps you could do something like initialise the 2 new token sets based on the shared data, then use existing chat logs to train it to understand the difference between input and output content? That's only a single extra phase.

New comment by __alexs in "Claude mixes up who said what"

__alexs — Thu, 09 Apr 2026 10:10:54 +0000

Why are tokens not coloured? Would there just be too many params if we double the token count so the model could always tell input tokens from output tokens?

New comment by __alexs in "The Future of Everything Is Lies, I Guess"

__alexs — Wed, 08 Apr 2026 16:51:53 +0000

Solving arbitrary logical problems seems to be equivalent to solving the halting problem so you are probably wise not to make that bet.

New comment by __alexs in "Every GPU That Mattered"

__alexs — Tue, 07 Apr 2026 09:07:53 +0000

A lot of GPUs in this list are basically just previous GPU but faster or more RAM. I kind of thought it was going to focus on interesting new architecture innovations.

New comment by __alexs in "Sam Altman may control our future – can he be trusted?"

__alexs — Tue, 07 Apr 2026 08:42:24 +0000

If you know even the basics of ethics then such claims are clearly nonsense. There is no stable context independent ethical behaviour. This is a great example of the dangers of motivated reasoning.

New comment by __alexs in "4D Doom"

__alexs — Wed, 01 Apr 2026 10:27:35 +0000

There have been other 4D video games with actually good performance though.

New comment by __alexs in "Claude Code's source code has been leaked via a map file in their NPM registry"

__alexs — Wed, 01 Apr 2026 10:24:27 +0000

No it isn't? Are you an AI?

New comment by __alexs in "Claude Code Unpacked : A visual guide"

__alexs — Wed, 01 Apr 2026 10:22:49 +0000

Many people seem to believe the Claude Code has some sort of secret sauce in the agent itself for some reason.

I have no idea why because in my experience Claude Code and the same models inside of Cursor behave almost identically. I think all the secret sauce is in the RLHF.

New comment by __alexs in "OpenAI closes funding round at an $852B valuation"

__alexs — Tue, 31 Mar 2026 21:48:34 +0000

I thought they needed $7 trillion or they'd be unable to keep training new models?

New comment by __alexs in "Claude Code's source code has been leaked via a map file in their NPM registry"

__alexs — Tue, 31 Mar 2026 13:56:56 +0000

Looking forward to someone patching it so that it works with non Anthropic models.

New comment by __alexs in "Claude Code's source code has been leaked via a map file in their NPM registry"

__alexs — Tue, 31 Mar 2026 13:42:43 +0000

Using some ML to derive a sentiment regex seems like a good actually?

New comment by __alexs in "Spring Boot Done Right: Lessons from a 400-Module Codebase"

__alexs — Mon, 30 Mar 2026 14:40:34 +0000

Not a strong indicator that Enterprises have good taste though is it?

New comment by __alexs in "Miscellanea: The War in Iran"

__alexs — Wed, 25 Mar 2026 17:39:02 +0000

The reset isn't the problem, the entirely nerfing the Red team is the problem. The US took steps to fail to learn from the exercise before it had even finished.

New comment by __alexs in "GitHub appears to be struggling with measly three nines availability"

__alexs — Mon, 23 Mar 2026 14:01:39 +0000

No they were slow at doing features before, and they are still slow afterwards.

New comment by __alexs in "Walmart: ChatGPT checkout converted 3x worse than website"

__alexs — Mon, 23 Mar 2026 13:07:02 +0000

It's not cynical it is materialism.

Shoppers do not want to pay to shop. Retailers pay thousands to encourage you to shop with them. They are the economic buyers of this feature.