Hacker News: JacobiX

LLM Works. Your Product Probably Doesn't

JacobiX — Thu, 30 Jul 2026 16:59:20 +0000

Article URL: https://saito.ai/llm-harness/

Comments URL: https://news.ycombinator.com/item?id=49112662

Points: 2

# Comments: 0

New comment by JacobiX in "Statement from Dario Amodei on our discussions with the Department of War"

JacobiX — Fri, 27 Feb 2026 09:34:58 +0000

>> We chose to forgo several hundred million dollars in revenue to cut off the use of Claude by firms linked to the Chinese Communist Party

You can’t choose to work with OFAC-designated entities.. there are very serious criminal penalties. Therefore, this statement is somewhat misleading in my opinion.

New comment by JacobiX in "Show HN: ShapedQL – A SQL engine for multi-stage ranking and RAG"

JacobiX — Thu, 29 Jan 2026 14:59:56 +0000

>> Apply diversity logic so your Agent/User doesn’t see 10 nearly identical results

On Instagram this is a good thing, but here the example is hotel and flight search, where a more deterministic result is preferable.

In the retrieve → filter stage, using predicate pushdown may be more performant: first filter using hard constraints, then apply hybrid search ?

New comment by JacobiX in "Is the golden age of Indie software over?"

JacobiX — Mon, 22 Dec 2025 22:40:03 +0000

>> it is getting harder for small software vendors

I think maybe this trend will continue and not specifically for indie developers, but for all software vendors. If AI becomes capable of producing genuinely highquality software, competition will intensify, and the industry will start to resemble the music industry. Alternatively, AI may continue to generate software that is not necessarily high quality but is largely indistinguishable from competing products; in that case, the market for lemons dynamic will apply. In either scenario, the value of software will decline...

New comment by JacobiX in "DeepSeekMath-V2: Towards Self-Verifiable Mathematical Reasoning [pdf]"

JacobiX — Thu, 27 Nov 2025 21:12:20 +0000

I think that mathematical proofs, as they are actually written, rely on natural language and on a large amount of implicit shared knowledge. They are not formalized in the Principia Mathematica sense, and they are even further from the syntax required by modern theorem provers. Even the most rigorous proofs such as those in Bourbaki are not directly translatable into a fully formal system.

New comment by JacobiX in "Technical Deflation"

JacobiX — Thu, 27 Nov 2025 15:43:07 +0000

In the end, the article says:

> writing functioning application code has grown easier thanks to AI.

> It's getting easier and easier for startups to do stuff.

> Another answer might be to use the fact that software is becoming free and disposable to your advantage.

For me, the logical conclusion here is: don't build a software startup!

New comment by JacobiX in "Gemini 3"

JacobiX — Tue, 18 Nov 2025 19:03:28 +0000

Yes; they nailed the root case but the implementation is not 100% correct

New comment by JacobiX in "Gemini 3"

JacobiX — Tue, 18 Nov 2025 18:42:29 +0000

Tested it on a bug that Claude and ChatGPT Pro struggled with, it nailed it, but only solved it partially (it was about matching data using a bipartite graph). Another task was optimizing a complex SQL script: the deep-thinking mode provided a genuinely nuanced approach using indexes and rewriting parts of the query. ChatGPT Pro had identified more or less the same issues. For frontend development, I think it’s obvious that it’s more powerful than Claude Code, at least in my tests, the UIs it produces are just better. For backend development, it’s good, but I noticed that in Java specifically, it often outputs code that doesn’t compile on the first try, unlike Claude.

New comment by JacobiX in "Disrupting the first reported AI-orchestrated cyber espionage campaign"

JacobiX — Thu, 13 Nov 2025 22:00:21 +0000

I have the feeling that we are still in the early stages of AI adoption, where regulation hasnt fully caught up yet. I can imagine a future where LLMs sit behind KYC identification and automatically report any suspicious user activity to the authorities... I just hope we won’t someday look back on this period with nostalgia :)

New comment by JacobiX in "Why JPEGs still rule the web (2024)"

JacobiX — Tue, 17 Jun 2025 16:14:53 +0000

MozJPEG, Guetzli and also Jpegli

New comment by JacobiX in "Why JPEGs still rule the web (2024)"

JacobiX — Tue, 17 Jun 2025 15:47:50 +0000

I loved the article, but it overlooks one important point: although the JPEG format is frozen, encoders are still evolving ! Advances such as smarter quantization, better perceptual models, and higher-precision maths enables us achieve higher compression ratios while sticking to a format that's supported everywhere :)

New comment by JacobiX in "Vacheron Constantin breaks the world record for most complicated wristwatch"

JacobiX — Sat, 12 Apr 2025 10:24:35 +0000

What I like about mechanical watches is that, having survived a near-death experience when quartz watches were introduced, they’ve evolved into a completely different kind of product. It’s fascinating that, unlike most other businesses and products, people don’t buy them for their utility, and the less automated their production process, the better. Brands like A. Lange & Söhne even pride themselves on assembling their movements twice.

When inefficiency and craftsmanship are considered features rather than flaws, you have an industry that won’t easily be replaced by AI or robots.

New comment by JacobiX in "Apache ECharts"

JacobiX — Tue, 08 Apr 2025 17:46:23 +0000

In a quick web demo, this library was the only one that could handle interactive viewing and manipulation of a very large graph using its GraphGL component ! I don’t think it's a well-known visualization library, but it's quite interesting ...

New comment by JacobiX in "A university president makes a case against cowardice"

JacobiX — Thu, 03 Apr 2025 13:58:45 +0000

Not sure if Michael Roth is related to Philip Roth, but it somehow reminds me of American Pastoral and that era of protests against the Vietnam War and its aftermath. I'm not entirely sure how those demonstrations compare to the ones we’re seeing today, but the parallels are striking

New comment by JacobiX in "Benchmarking RSA Key Generation"

JacobiX — Fri, 03 Jan 2025 16:26:33 +0000

It is fascinating (to me at least) that almost all RSA implementations rely on a probabilistic primality test, even though the probability of picking a pseudoprime is extremely small.

New comment by JacobiX in "Show HN: Squey, an open-source GPU-accelerated data visualization software"

JacobiX — Tue, 08 Oct 2024 17:12:46 +0000

Thank you, I'll look into it more closely, fortunately it builds using CMake / Clang, and cross-platform libs ... might be possible to port it to MacOS after some tweaks.

New comment by JacobiX in "Show HN: Squey, an open-source GPU-accelerated data visualization software"

JacobiX — Tue, 08 Oct 2024 16:25:20 +0000

Impressive project, judging by commits and features, it's clear that significant effort has been poured into this :) Unfortunately, there's no specific MacOS installation method provided, unsure if buildable from source ?

New comment by JacobiX in "Ask HN: What's the "best" book you've ever read?"

JacobiX — Mon, 07 Oct 2024 16:59:14 +0000

I second this... to this day, I am in awe! Such a profound and readable book in the same time. Each character lives within me, and in many situations, one of them emerges :)

As Nabokov said, it is the supreme masterpiece of 19th-century prose.

New comment by JacobiX in "DuckDB Isn't Just Fast"

JacobiX — Tue, 11 Jun 2024 10:38:21 +0000

In my experience, recent versions work well for out-of-core computation (I use it frequently). What is the size of the Parquet file that you are using? - I can't find the details in you previous comments.

New comment by JacobiX in "GPT-4.5 or GPT-5 being tested on LMSYS?"

JacobiX — Mon, 29 Apr 2024 17:43:53 +0000

The model provides verbose answers even when I asked for more succinct ones. It still struggles with arithmetic (for example, it incorrectly stated "7739 % 23 = 339 exactly, making 23 a divisor"). When tested with questions in French, the responses were very similar to those of GPT-4. It is far better in knowledge based questions, I've asked this difficult one (it is not 100% correct but better than other LLMs) :

In Anna Karenina what does it mean: most of us prefer the company of Claras ?