HNHacker News
TopNewBestAskShowJobs

nopinsight

7,312 karma · joined March 14, 2009

Co-Founder of a stealth startup

Working on:

agentic systems, natural language understanding, machine learning

Comments here do not represent any companies or organizations I am associated with.

Email: My username at the best known email service beginning with g.

submissionscomments
nopinsight··on Why does the U.S. always run a trade deficit?
>> ...median urban household income...

You seem to miss the word 'disposable' when you cited the chart and table in [0] you linked to.

China's nominal GDP per capita is already on par with Malaysia's.

https://en.wikipedia.org/wiki/List_of_countries_by_GDP_(nomi...

Does your analysis on the relative standards of living of China's and Malaysia's populations still stand?

It's been more than 10 years since I last visited China. I've been to Malaysia a couple of times not too long ago. From what I see from random social media clips of regular people, their urban standards of living do not differ much.

nopinsight··on Thoughts on thinking
Au contraire, reasoning LLMs free me from mundane cognitive tasks, e.g. selecting websites to read from among search results, assessing level of evidence in papers before reading them in more detail, routine coding, and many other repetitive tasks. It allows me to spend mental effort on more abstract and strategic thoughts.

We should seize the opportunity to elevate our mental skills to a higher plane, aided by these “cognitive elves”.

nopinsight··on A Research Preview of Codex
Training data, esp interaction data from agentic coding tools, are important for that. See also: Windsurf acquisition.
nopinsight··on A Research Preview of Codex
There’s a recent article with experiments suggesting LLMs are better at bug fixing than coding, iirc. It’s from a company with a relevant product though.
nopinsight··on A Research Preview of Codex
With Agentic RL training and sufficient data, AI operating at the level of average senior engineers should become plausible in a couple to a few years.

Top-tier engineers who integrate a deep understanding of business and user needs into technical design will likely be safe until we get full-fledged AGI.

nopinsight··on In 2025, venture capital can't pretend everything is fine any more
> 4.5/o3 doesn't seem hugely more intelligent then 3.0

I disagree with 3.0, but perhaps that feels true for 4.0 or even 3.5 for some queries.

The reason is that when LLMs are asked questions whose answers can be interpolated or retrieved from their training data, they will likely use widely accepted human knowledge or patterns to compose their responses. (This is a simplification of how LLMs work, just to illustrate the key point here.) This knowledge has been refined and has evolved through decades of human experiments and experiences.

Domain experts of varying intelligence will likely come up with similar replies on these largely routine questions as well.

The difference shows up when you pose a query that demands deep reasoning or requires expertise in multiple fields. Then, frontier reasoning models like o3 can sometimes form creative solutions that are not textbook answers.

I strongly suspect that Reinforcement Learning with feedback from high-quality simulations or real environments will be key for these models' capabilities to surpass those of human experts.

Superhuman milestones, equivalent to those achieved by AlphaGo and AlphaZero between 2016 and 2018, might be reached in several fields over the coming years. This will likely happen first in fields with rapid feedback loops and highly accurate simulators, e.g. math problem solving (as opposed to novel mathematical research), coding (as opposed to product innovation).

nopinsight··on Show HN: I’m 16 years old and working on my first startup, a study app
She appears to be a very good product developer. However, unless the hypothetical VC knows more about their upcoming plan, it doesn't look like a fundable business with a strong moat and the potential to become a unicorn or larger business.
nopinsight··on Gemini 2.5 Pro Preview
Livebench.ai actually suggests the new version is better on most things.

https://livebench.ai/#/

nopinsight··on Qwen3: Think deeper, act faster
Current models are quite far away from human-level physical reasoning (paper below). An upcoming version of models trained on world simulation will probably do much better.

PHYBench: Holistic Evaluation of Physical Perception and Reasoning in Large Language Models

https://phybench-official.github.io/phybench-demo/

nopinsight··on AIs Are Disseminating Expert-Level Virology Skills
"Can AI meaningfully help with bioweapons creation? On our new Virology Capabilities Test (VCT), frontier LLMs display the expert-level tacit knowledge needed to troubleshoot wet lab protocols."

OpenAI’s o3 now outperforms 94% of expert virologists." -- thread by a co-author, https://x.com/DanHendrycks/status/1914696657813561799

Paper: Virology Capabilities Test (VCT): A Multimodal Virology Q&A Benchmark

https://www.virologytest.ai/vct_paper.pdf

nopinsight··on Forecaster reacts: METR's bombshell paper about AI acceleration
> software engineering skills are easier to train than other skills.

I think the author meant it's easier to train (reasoning) LLM models on [coding] skills than most other tasks. I agree on that. Data abundance, near-immediate feedback, and near-perfect simulators are why we've seen such rapid progress on most coding benchmarks so far.

I'm not sure if he included high-level software engineering skills such as designing the right software architecture for a given set of user requirements in that statement.

---

For humans, I think the fundamentals of coding are very natural and easy for people with certain mental traits, although that's obviously not the norm (which explains the high wages for some software engineers).

Coding on large, practical software systems is indeed much more complex with all the inherent and accidental complexity. The latter helps explain why AI agents for software engineering will require some human involvement until we actually reach full-fledged AGI.

nopinsight··on LLM-powered tools amplify developer capabilities rather than replacing them
The SWE-bench verified score for frontier LLMs will probably reach/surpass 90% by the end of the year.

Agentic AI will learn to complete a larger and larger chunk of the practical software development process without much human input.

nopinsight··on LLM-powered tools amplify developer capabilities rather than replacing them
One way to think of this is as the Baumol effect* within software development.

Expert humans are still quite a bit better than LLMs at nuanced requirements understanding and architectural design for now. Actual coding will increasingly become a smaller and cheaper part of the process, while the parts where human input cannot be reduced as much will take up a larger proportion of time and cost.

* Not everything here applies, but many will be. https://en.m.wikipedia.org/wiki/Baumol_effect

nopinsight··on Olympic Gold Medalist Dominates the 100M at Her Kid's Sports Day Event
https://news.ycombinator.com/item?id=43743198
nopinsight··on Olympic Gold Medalist Dominates the 100M at Her Kid's Sports Day Event
Verbatim quote: “The difference between me and an average Joe in strength technique and speed was still less than the difference between me and the Olympic level athletes that I occasionally competed (or worse, trained) with.”

I’m not disputing the gaps in technique, just to be clear.

nopinsight··on Olympic Gold Medalist Dominates the 100M at Her Kid's Sports Day Event
I'm curious. Do you mean it figuratively? I ask Claude Sonnet 3.7 Extended Thinking since it's usually reliable and the stats for running strongly suggest that for most competitions, the top percentile is closer to world-class athletes than the average person to top percentile athletes (possibly except Marathon).

*100m Sprint*:

1st percentile: ~18-20 seconds

50th percentile: ~14-15 seconds

99th percentile: ~11-11.5 seconds

Elite world-class: ~9.8-10.2 seconds

World record: 9.58 seconds (Usain Bolt, 2009)

*1-Mile Run*:

1st percentile: ~12-15 minutes

50th percentile: ~8-9 minutes

99th percentile: ~4:30-5:00 minutes

Elite world-class: ~3:45-3:55

World record: 3:43.13 (Hicham El Guerrouj, 1999)

*Marathon* (26.2 miles):

1st percentile: ~6+ hours

50th percentile: ~4:30-5:00 hours

99th percentile: ~3:00-3:15 hours

Elite world-class: ~2:05-2:10

World record: 2:00:35 (Kelvin Kiptum, 2023)*

nopinsight··on OpenAI's new reasoning AI models hallucinate more
The reward it gets from the reinforcement learning (RL) process probably didn’t include a sufficiently strong weight on being truthful.

Reward engineering for RL might be the most important area of research in AI now.

nopinsight··on OpenAI looked at buying Cursor creator before turning to Windsurf
Many of them likely won’t switch immediately. They could also try to keep them with sweet offers, like generous usage quotas, early access to the latest models, etc.

Once sufficient data is gathered, the next generation models will be among the very best at agentic coding, which leads to stronger stickiness, and so on.

nopinsight··on OpenAI looked at buying Cursor creator before turning to Windsurf
OpenAI‘s early investment in Cursor was a masterstroke. Acquiring Windsurf would be another.

Next advances in coding AI depend on real-world coding data, esp how professional developers use agentic AI for coding + other tasks.

RL works well on sufficiently large base models as shown by rapid progress on verifiable problems with good training data, e.g. competition math, competitive coding problems, scientific question answering.

Training LLMs on detailed interaction data from AI-powered IDEs could become a powerful flywheel leading to the automation of practical coding.

nopinsight··on OpenAI o3 and o4-mini
A standard term people use for what you describe is superintelligence, not AGI.

Current frontier models are better than average humans in many skills but worse in others. Ethan Mollick calls it “jagged frontier” which sounds about right.

nopinsight··on OpenAI o3 and o4-mini
Research by METR suggests that frontier LLMs can perform software tasks over exponentially longer time horizon required for human engineers, with ~7-month for each doubling. o3 is above the trend line.

https://x.com/METR_Evals/status/1912594122176958939

—-

The AlexNet paper which kickstarted the deep learning era in 2012 was ahead of the 2nd-best entry by 11%. Many published AI papers then advanced SOTA by just a couple percentage points.

o3 high is about 9% ahead of o1 high on livebench.ai and there are also quite a few testimonials of their differences.

Yes, AlexNet made major strides in other aspects as well but it’s been just 7 months since o1-preview, the first publicly available reasoning model, which is a seminal advance beyond previous LLMs.

It seems some people have become desensitized to how rapidly things are moving in AI, despite its largely unprecedented pace of progress.

Ref:

- https://proceedings.neurips.cc/paper_files/paper/2012/file/c...

- https://livebench.ai/#/

nopinsight··on What can we take away from the ‘stochastic parrot’ saga?
The fact that much of physics can be so elegantly described by math suggests the structures of our math could be quite universal, at least in our universe.

Check out the problems in the MATH dataset, especially Level 5 problems. They are fairly advanced (by most people’s standards) and most are not dependent on which N in the base-N system used to solve them. The answers would be different of course but the structures of the problems and solutions remain largely intact.

Website for tracking IQ measurements of LLMs:

https://www.trackingai.org/

The best one already scores higher than all but the top 10-20% of most populations.

nopinsight··on What can we take away from the ‘stochastic parrot’ saga?
My request:

“Could you explain this sentence concisely?

For the skeptics: Scoring just 10% or so in Math-Perturb-Hard below the original MATH Level 5 (hardest) dataset seems in line with or actually better than most people would do.”

Gemini 2.5 Pro:

“The sentence argues that even if a model's score drops by about 10% on the "Math-Perturb-Hard" dataset compared to the original "MATH Level 5" (hardest) dataset, this is actually a reasonable, perhaps even good, outcome. It suggests this performance decrease is likely similar to or better than how most humans would perform when facing such modified, difficult math problems.”

nopinsight··on What can we take away from the ‘stochastic parrot’ saga?
Please check out the post on Math-Perturb-Hard conveniently linked to above before making a comment without responding to it.

A relevant bit:

“for MATH-P-Hard, we make hard perturbations, i.e., small but fundamental modifications to the problem so that the modified problem cannot be solved using the same method as the original problem. Instead, it requires deeper math understanding and harder problem-solving skills.”

nopinsight··on What can we take away from the ‘stochastic parrot’ saga?
Solving hard math problems requires understanding the structure of complex mathematical reasoning. No animal is known to be capable of that.

Most definitions and measurements of intelligence by most laypeople and psychologists include the ability to reason, with mathematical reasoning widely accepted as part of or a proxy for it. They are imperfect but “intelligence” does not have a universally accepted definition.

Do you have a better measurement or definition?

nopinsight··on What can we take away from the ‘stochastic parrot’ saga?
For the skeptics: Scoring just 10% or so in Math-Perturb-Hard below the original MATH Level 5 (hardest) dataset seems in line with or actually better than most people would do.

Does that mean most people are merely parrots too?

https://math-perturb.github.io/

https://arxiv.org/abs/2502.06453

Leaderboard: https://math-perturb.github.io/#leaderboard

Anyone who continues to use the parrot metaphor should support it with evidence at least as strong as the “On the Biology of a Large Language Model” research by Anthropic which the article refers to:

https://transformer-circuits.pub/2025/attribution-graphs/bio...

nopinsight··on AI 2027
If by “several” orders of magnitude, you mean 3-5, then we might be there by 2030 or earlier.

https://situational-awareness.ai/from-gpt-4-to-agi/

nopinsight··on I genuinely don't understand why some people are still bullish about LLMs
>> unaided humans cannot operate successfully in many situations

> Absolute nonsense driven by a total lack of historical perspective or knowledge.

An LLM can give you a list of examples:

Historical Examples:

- During historical epidemics, structured record-keeping and statistical analysis (such as John Snow’s cholera maps in 1854) significantly improved outcomes.

- Development of physics, architecture, and engineering depended heavily on tools such as abacus, logarithmic tables, calculators, slide rules, to supplement human cognitive limitations.

- Astronomical calculations in ancient civilizations (Babylonian, Greek, Mayan) depended heavily on abacuses, tables, and other computational tools.

- The pyramids in ancient Egypt required extensive use of tools, mathematics, coordinated human labor, and sophisticated organization.

nopinsight··on I genuinely don't understand why some people are still bullish about LLMs
Can our brains recall precise citations to tens of papers we read a while ago? For the vast majority, no. LLMs function somewhat similarly to our brains in many ways, as opposed to classical computers.

Their strengths and flaws differ from our brains, to be sure, but some of these flaws are being mitigated and improved on by the month. Similarly, unaided humans cannot operate successfully in many situations. We build tools, teams, and institutions to help us deal with them.

nopinsight··on Stargate Project: SoftBank, OpenAI, Oracle, MGX to build data centers
1) It’s an example of a domain an LLM can do better than humans. A 2012 system was not able to do myriad other things LLMs can do and thus not qualified as general intelligence.

2) As mentioned in the chart label, earlier systems require manual symptom extraction.

3) An important point well articulated by a cancer genomics faculty member at Harvard:

“….Now, back to today: The newest generation of generative deep learning models (genAI) is different.

For cancer data, the reason these models hold so much potential is exactly the reason why they were not preferred in the first place: they make almost no explicit data assumptions.

These models are excellent at learning whatever implicit distribution from the data they are trained on

Such distributions don’t need to be explainable. Nor do they even need to be specified

When presented with tons of data, these models can just learn, internalize & understand…..”

More here: https://x.com/simocristea/status/1881927022852870372?s=61&t=...

← PreviousPage 3 of 34Next →