Hacker News: irthomasthomas

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

irthomasthomas — Thu, 09 Apr 2026 12:59:57 +0000

I have suffered a lot with this recently. I have been using llms to analyze my llm history. It frequently gets confused and responds to prompts in the data. In one case I woke up to find that it had fixed numerous bugs in a project I abandoned years ago.

New comment by irthomasthomas in "Project Glasswing: Securing critical software for the AI era"

irthomasthomas — Wed, 08 Apr 2026 00:04:55 +0000

That still leaves open the possibility that they reduce model quality due to profit. ;p

New comment by irthomasthomas in "Trump says 'a whole civilization will die tonight' if Iran does not make a deal"

irthomasthomas — Tue, 07 Apr 2026 19:53:17 +0000

Apparently they came with 75 meters of bombing an nuclear power plant already. A plant with 10x the material Chernobyl had, and in vulnerable above-ground storage. https://www.ndtv.com/world-news/middle-east-war-why-attacks-...

New comment by irthomasthomas in "Artemis computer running two instances of MS outlook; they can't figure out why"

irthomasthomas — Thu, 02 Apr 2026 21:34:35 +0000

Is this incident not reason enough? Astronauts in space are needing remote support to debug it, and taking up priceless mission time.

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

irthomasthomas — Tue, 31 Mar 2026 17:27:00 +0000

Actually, this could be a case where its useful. Even it only catches half the complaints, that's still a lot of data, far more than ordinary telemetry used to collect.

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

irthomasthomas — Tue, 31 Mar 2026 14:13:12 +0000

This just proves its vibe coded because LLMs love writing solutions like that. I probably have a hundred examples just like it in my history.

New comment by irthomasthomas in "TurboQuant: Redefining AI efficiency with extreme compression"

irthomasthomas — Wed, 25 Mar 2026 17:58:09 +0000

it enables models larger than was previously possible.

New comment by irthomasthomas in "TurboQuant: Redefining AI efficiency with extreme compression"

irthomasthomas — Wed, 25 Mar 2026 12:55:18 +0000

Efficiency gains can be used to make existing models more profitable, or to make new larger and more intelligent models.

New comment by irthomasthomas in "Vatican Rebukes Peter Thiel's Antichrist Lectures in Rome"

irthomasthomas — Sun, 22 Mar 2026 14:55:40 +0000

Its possibly just an SEO trick. People have been calling Thiel the antichrist for a long time.

New comment by irthomasthomas in "Profiling Hacker News users based on their comments"

irthomasthomas — Sun, 22 Mar 2026 10:57:57 +0000

A friend made a cli tool, ideal for agents, which does this and can aggregate intelligence across multiple platforms.

https://github.com/bm-github/owasp-social-osint-agent

New comment by irthomasthomas in "Many SWE-bench-Passing PRs would not be merged"

irthomasthomas — Thu, 12 Mar 2026 15:45:41 +0000

Have you tried meta-prompts e.g. "Rewrite the prompt to improve the perceived taste and expertise of the author"

New comment by irthomasthomas in "No, it doesn't cost Anthropic $5k per Claude Code user"

irthomasthomas — Tue, 10 Mar 2026 10:50:05 +0000

Opus doubled in speed with version 4.5, leading me to speculate that they had promoted a sonnet size model. The new faster opus was the same speed as Gemini 3 flash running on the same TPUs. I think anthropics margins are probably the highest in the industry, but they have to chop that up with google by renting their TPUs.

New comment by irthomasthomas in "Pentagon formally labels Anthropic supply-chain risk"

irthomasthomas — Thu, 05 Mar 2026 22:10:53 +0000

They will rename it The Free Democratic Republic of America.

New comment by irthomasthomas in "Agentic Engineering Patterns"

irthomasthomas — Wed, 04 Mar 2026 12:44:50 +0000

Here is an example where the prompt was only a few hundred tokens and the output reasoning chain was correct, but the actual function call was wrong https://x.com/xundecidability/status/2005647216741105962?s=2...

New comment by irthomasthomas in "Agentic Engineering Patterns"

irthomasthomas — Wed, 04 Mar 2026 12:43:56 +0000

New comment by irthomasthomas in "Bet on German Train Delays"

irthomasthomas — Wed, 04 Mar 2026 12:24:20 +0000

No they don't, the practice was banned some time ago. You now require a "insurable interest". https://en.wikipedia.org/wiki/Marine_Insurance_Act_1745#:~:t...

New comment by irthomasthomas in "Bet on German Train Delays"

irthomasthomas — Wed, 04 Mar 2026 11:57:46 +0000

People used to bet on ships sinking and sailors drowning. Till they learned better.

Edit: This was common until Parliament passed the Marine Insurance Act of 1745.

Before that, speculators could take out "wagering policies" on vessels they had no connection to. This created "coffin ships" - unseaworthy vessels sent to sea because the insurance payout for a wreck was worth more than the ship itself. The law introduced "insurable interest," meaning you cannot bet on a disaster unless you stand to lose something if it happens. This removed the incentive for sabotage and murder for profit.

Modern prediction markets are heading toward the same problem. Betting on train delays or bridge collapses without having any stake gives bad actors a reason to cause it. If the cost of sabotage is lower than the payout, the market effectively pays for the disaster to happen.

Whoever downvoted this wants you to ignore centuries of legal precedent designed to prevent exactly this kind of blood money. Those who ignore the lessons of the past learn wisdom in blood... https://en.wikipedia.org/wiki/Coffin_ship_(insurance)#:~:tex... https://en.wikipedia.org/wiki/Marine_Insurance_Act_1745#:~:t...

New comment by irthomasthomas in "Agentic Engineering Patterns"

irthomasthomas — Wed, 04 Mar 2026 11:50:00 +0000

I do this too, but then you need some method to handle it, because now you have to read and test and verify multiple work streams. It can become overwhelming. In the past week I had the following problems from parallel agents:

Gemini running an benchmark- everything ran smoothly for an hour. But on verification it had hallucinated the model used for judging, invalidating the whole run.

Another task used Opus and I manually specified the model to use. It still used the wrong model.

This type of hallucination has happened to me at least 4-5 times in the past fortnight using opus 4.6 and gemini-3.1-pro. GLM-5 does not seem to hallucinate so much.

So if you are not actively monitoring your agent and making the corrections, you need something else that is.

New comment by irthomasthomas in "Why XML tags are so fundamental to Claude"

irthomasthomas — Mon, 02 Mar 2026 10:40:20 +0000

The main thing i use xml tags for is seperating content from instructions. Say I am doing prompt engineering, so that the content being operated on is itself a prompt then I wrap it with

draft prompt

instructions for modifying draft prompt

If I don't do this, a significant number of times it responds to the instructions in the draft.

New comment by irthomasthomas in "Our Agreement with the Department of War"

irthomasthomas — Sun, 01 Mar 2026 13:33:50 +0000

Why would you want a duplicitous CEO in charge of your countries terminator systems?