HNHacker News
TopNewBestAskShowJobs

christianqchung

360 karma · joined December 21, 2023

submissionscomments
christianqchung··on Open models by OpenAI
> Training: The gpt-oss models trained on NVIDIA H100 GPUs using the PyTorch framework [17] with expert-optimized Triton [18] kernels2. The training run for gpt-oss-120b required 2.1 million H100-hours to complete, with gpt-oss-20b needing almost 10x fewer.

This makes DeepSeek's very cheap claim on compute cost for r1 seem reasonable. Assuming $2/hr for h100, it's really not that much money compared to the $60-100M estimates for GPT 4, which people speculate as a MoE 1.8T model, something in the range of 200B active last I heard.

christianqchung··on OpenAI raises $8.3B at $300B valuation
Facebook users do not know Llama anywhere near the degree the X network does Grok, and both are pallid in comparison to Chatgpt.
christianqchung··on OpenAI raises $8.3B at $300B valuation
I thought they were going for $340B?[0]

[0] https://www.cnbc.com/2025/01/30/openai-in-talks-to-raise-up-...

christianqchung··on Study mode
> This isn't a knock on your culture: the US is at a point where a stabucks barista part-time is paid more than some schoolteachers.

I don't think this is meaningfully true. I found a resource that shows the average teacher salary to be $72,030 [0]. The average starting salary is lower at $46,526, but a 40 hour workweek at $20 for a Starbucks barista tips-included is about $41k. Here in Massachusetts, the average teacher salary is $92,076. In Mississippi, it's $53,704. You can maybe find some full time (not part time) Starbucks baristas that make slightly more than starting teachers, but after a couple of years the teacher will pull ahead. However, since the higher paying Starbucks jobs are in places with higher costs of living, I would assume that the teacher pay would be higher in those places too, so it's a wash.

> "We've tried nothing and run out of ideas!", as a famous American saying.

Ironically Mississippi of all states has experimented by holding back more poor performing kids instead of letting them advance to the next grade, with some success in rising test scores: "Boston University researchers released a study this year comparing Mississippi students who were narrowly promoted to fourth grade to those who just missed the cutoff. It found that by sixth grade, those retained had substantial gains on English language arts scores compared with those who were promoted, especially among African-American and Hispanic students." [1].

This doesn't disprove what you're saying (and there are some caveats to the Mississippi experiment), but there is definitely low hanging fruit to improve the American teaching system. Just because teaching is a thousands year old profession doesn't mean modern day processes can't be improved by ways not involving salaries/direct training.

[0] https://www.nea.org/resource-library/educator-pay-and-studen...

[1] https://www.wsj.com/us-news/education/more-states-threaten-t...?

christianqchung··on Epic Recall, Epic Fail
The failed recall vote was trying to solve political gridlock for the DPP by winning enough seats for a legislative majority. My point is that in a well functioning system, this shouldn't be a step the DPP should take if they could get good results by compromising with the KMT/TPP. Having said that, I think the KMT/TPP are poorly aligned for Taiwan's interests, so this makes the situation extremely difficult.
christianqchung··on Epic Recall, Epic Fail
> You know what would make the people happy? If they just got to governing instead of grandstanding and brawling and actually tried to form a consensus instead of shouting each other down and trying to achieve absolute power. I mean, isn’t that what democracy is supposed to be about?

Kind of hard to do that when leadership doesn't control the legislature. See the Biden years after losing the House in 2022, Trump after losing the House in 2018, Obama after losing the house in 2010, Bush losing the House in 2006, and so on. Why does legislative compromise look like it comes easier in some countries than other? Not a snarky remark, an open question. German conservatives routinely coalition with social democrats, vice versa in Denmark, Switzerland has had a permanent grand coalition for a long while now, etc.

I'd like to add that existentialism in Taiwan is entirely warranted, and that denialism about an imminent military invasion was also widespread in Ukraine right up until Febuary 2022. That doesn't excuse poor domestic governance (though poor is likely a stretch, objectively the performance of the economy is very high, currently the PPP/capita is about the same as Denmark), but it does explain the priorities in messaging.

christianqchung··on Qwen3-235B-A22B-Thinking-2507
That's true, I certainly wouldn't be able to tell. I was thinking on the order of a 20% score vs 70%, but I realize that's not a very compelling range for my point when people are boasting about <5% shifts.
christianqchung··on Qwen3-235B-A22B-Thinking-2507
They made it very clear that they were reporting that score for the non-thinking model[0]. I still don't have any guesses as to what happened here, maybe something format related. I can't see a motivation to blatantly lie on a benchmark which would very obviously be publicly corrected.

[0] https://x.com/JustinLin610/status/1947836526853034403

christianqchung··on Qwen3-235B-A22B-Thinking-2507
There is something of a middle ground here for benchmark skepticism. Big companies wouldn't want a massive divergence between benchmarks and real performance that people could actually notice, and I'd argue for the most part that this hasn't happened too much (although above I posted a problem with Qwen and ARC). However, finetunes by random people/groups don't carry the same downside so I'm basically skeptical of all finetunes before using them for a particular case.
christianqchung··on Qwen3-235B-A22B-Thinking-2507
For what it's worth, the Qwen team misreported an ARC-AGI score benchmark on the non-thinking model by a factor of 4, which has not been explained yet. They claimed a score of 41.8% on ARC-AGI 1 [0] which is much higher than what non-chain of thought models have been able to achieve (GPT 4.5 got 10%). The ARC team later benchmarked it at 11%[1], which is still a high score, but not the same as 41.8%. It's still probably a significant update on the model though.

[0] https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507

[1] https://x.com/arcprize/status/1948453132184494471

christianqchung··on AI comes up with bizarre physics experiments, but they work
The thread is full of people debating what AI means and whether ML/optimization algorithms count as AI. Laypeople don't think of machine learning when they see AI, they think of chatbots. I would argue even for a techy magazine this is a bad term to use without spending two sentences to clarify the distinction.

Examples of people being confused:

rlt: The discovering itself doesn’t seem like the interesting part. If the discovery wasn’t in the training data then it’s a sign AI can produce novel scientific research / experiments.

wizzwizz4 in reply: It's not that kind of AI. We know that these algorithms can produce novel solutions. See https://arxiv.org/abs/2312.04258, specifically "Urania".

About a quarter of the comments here I just have to assume what definition of AI they're talking about, which changes the meaning and context significantly.

christianqchung··on The current hype around autonomous agents, and what actually works in production
Weird deflection. GPT 3, a 175 billion parameter model, came out in 2020, so it very much is not before LLMs. I guess you can say the story happened before ChatGPT by a few weeks. As for putting lost in quotation marks, I have no idea why you think the quantity of money is relevant to the outcome. No one was expecting them to file bankruptcy over this, only to follow the original agreement.
christianqchung··on My Family and the Flood
What an insensitive, assumptive, stupid remark. You can't possibly know that the person you replied to behaves as you claim. It's 2025, the firebombing of Tokyo is widely recognized now, maybe not by most normies but certainly by any historical adjacent nerd.
christianqchung··on Seven Engineers Suspended After $2.3M Bridge Includes 90-Degree Turn
Based on the use of "Mum", very probably.
christianqchung··on The Real GenAI Issue
> the biggest polluters and criminals in the climate story are still by far the fossil fuel industry (eg. Shell, Exxon)

What does this actually mean? It feels like when people say this it implies that gas companies just burn gas randomly for no reason. They sell their gas to everyday consumers, actual people. Why is this Shell/Exxon's fault and not consumers?

christianqchung··on How large are large language models?
This is a bad article. Some of the information is wrong, and it's missing lots of context.

For example, it somehow merged Llama 4 Maverick's custom Arena chatbot version with Behemoth, falsely claiming that the former is stopping the latter from being released. It also claims 40B of internet text data is 10B tokens, which seems a little odd. Llama 405B was also trained on more than 15 trillion tokens[1], but the post claims only 3.67 trillion for some reason. It also doesn't mention Mistral large for some reason, even though it's the first good European 100B+ dense model.

>The MoE arch. enabled larger models to be trained and used by more people - people without access to thousands of interconnected GPUs

You still need thousands of GPUs to train a MoE model of any actual use. This is true for inference in the sense that it's faster I guess, but even that has caveats because MoE models are less powerful than dense models of the same size, though the trade-off has apparently been worth it in many cases. You also didn't need thousands of GPUs to do inference before, even for the largest models.

The conclusion is all over the place, and has lots of just weird and incorrect implications. The title is about how big LLMs are, why is there such a focus on token training count? Also no mention of quantized size. This is a bad AI slop article (whoops, turns out the author accidentally said it was AI generated, so it's a bad human slop article).

[1] https://ai.meta.com/blog/meta-llama-3-1/

christianqchung··on GCP Outage
Goomba fallacy
christianqchung··on Magistral — the first reasoning model by Mistral AI
I don't understand why the benchmark selections are so scattered and limited. It only compares Magistral Medium with Deepseek V3, R1, and the other close weighted Mistral Medium 3. Why did they leave off Magistral Small entirely, alongside comparisons with Alibaba Qwen or the mini versions of o3 and o4?
christianqchung··on AI Malware Is Here: New Report Shows How Fake AI Tools Are Spreading Ransomware
>And with such a size, you have more compute than OpenAI and can train a frontier misaligned model or modify an open source one

That doesn't seem like a reasonable conclusion based on Open AI and leading lab expenditures. They need training data too, would one group actually be able to amass all of this? If you hijack 30 million random computers that's not as nearly as useful as 300K B200 GPUS for model training. Am I missing something?

christianqchung··on AI video you can watch and interact with, in real-time
Is it possible that this behavior is a result from training on Google Maps or something similar? I tried to walk off a bridge and you get completely stuck, which is the only reason I can think of that, other than not having first person video views of people walking off bridges.
christianqchung··on Deepseek R1-0528
Quantization alone does not explain it. It's mostly custom hardware[0].

[0] https://groq.com/the-groq-lpu-explained/

christianqchung··on What makes a good engineer also makes a good engineering organization (2024)
"everything that isn't tiktok/podcasts is propaganda"
christianqchung··on Intellect-2 Release: The First 32B Model Trained Through Globally Distributed RL
Third party fine tuned open weighted LLMs tend to be good at a handful of benchmarks, but parity or lower on others compared to the original model. There are some exceptions like Nvidia's Nemotron series, but the differences generally are so small as to be imperceptible. Deepseek released finetunes of several Qwen and Llama models alongside R1, and while they were better in some select (mostly math) and coding domains, there were problems resulting from fine tuning that didn't result in them overtaking the original models in usage.
christianqchung··on O3 beats a master-level GeoGuessr player, even with fake EXIF data
Wouldn't that apply to any website with image hosting abilities though? Why does that apply to OpenAI in particular?

On an unrelated note, I like your blog.

christianqchung··on O3 beats a master-level GeoGuessr player, even with fake EXIF data
ChatGPT doesn't let you share chats with images (why?), but here is what o3 said for an image with memory turned off:

"Thought for 11 seconds

It’s hard to pin down an exact spot from this single photo, but a few visual clues narrow the possibilities:

Temperate-zone trees: Most of the trees have lost their leaves, but scattered white-pines or spruces remain green. That mix—deciduous hardwoods plus a sprinkling of conifers—is typical of the northeastern United States and southeastern Canada.

Broad, moderately fast river: The channel looks 75-100 m wide with swirling surface currents but no large rapids, suggesting a mid-sized river in its lower reaches—something like the Merrimack, Connecticut, Delaware, or Susquehanna rather than a small mountain stream.

Late-winter / early-spring light: Bare branches and a clear blue sky with fairly high sun angle hint at March–April in that latitude range.

So my best guess is that you’re standing on the bank of a northeastern U.S. river—perhaps somewhere in New England or the Mid-Atlantic—during late winter or early spring. If you can share any nearby landmarks, town names, or additional photos, I might be able to zero in more precisely."

With memory turned on:

"Thought for 5 seconds

Beautiful spot! From what I can see:

a broad, slow‑moving river with a sandy/rocky bank in the foreground

bare deciduous trees mixed with some white pines on the far shore (so it’s probably early spring or late autumn)

no mountains on the horizon, just gently rolling terrain typical of low‑lying floodplains in the U.S. Northeast

That mix of hardwoods and pines, plus the flat, braided water surface, reminds me a lot of the Merrimack or Concord Rivers in Massachusetts or southern New Hampshire—both have public footpaths where you can stand right at the water’s edge like this.

Of course, many temperate rivers look similar. If you can give me a hint (e.g., the town or nearby landmark), I can narrow it down further—maybe even point you to trail maps or history of the exact stretch you photographed."

Southern NH is correct, and I am certain it is drawing from memory/past chats. However, I can't replicate a specific behavior I once had, which is in temporary chat (no past chat/memory enabled), it said that it guessed where the photo was taken based on my location.

christianqchung··on The Friendship Recession: The lost art of connecting
The comment you replied to isn't about whether friendships are transactional. They detailed a long list of ways that your theoretical and "realistic" friendships are bad and unrealistic definitions. If you didn't have friends closer to your first definition in most ways than your second, you haven't had a close friend, which isn't something to be shamed over but is something to consider.
christianqchung··on Could GPT help with dating anxiety?
No. Betteridge's law. More seriously, while this piece is ancient in AI terms (May 2023), I don't think genuine emotional and social intelligence is something that can be learned at an average level by talking to AI or reading. Using the voice models is a step up from this, but I still think they're too tuned to following your instructions without nuance for something like this. If reading was enough to pick up social and emotional skills, I'd think that people who read the right books would be masters at several trades if it gave even 10% of the experience that real world practice did.

I'm also not trying to be reactionary and dismissive, but how are you supposed to learn social cues from AI right now? In an optimal case, the LLM predicts accurately what would happen. Maybe you could say something awkward and the AI would say back <s/he lets out an exasperated sigh and turns away>; but in real life you have to notice these cues among a barrage of other factors. Would this really help anyone who is this desperate? Additionally, the tone of how you say words matters almost as much as the content, and this is missing from text.

I concede that in extreme cases, some people could learn stuff from trying this, and that's a good thing. I just don't really know how much, who exactly, how, and whether they'd learn incorrect stuff as well.

christianqchung··on Kennedy promises to find environmental causes of autism
"Kennedy also claimed most autism cases are severe before citing a statistic from the report that said 25 percent of children with autism are nonverbal. HHS did not respond to a question as to whether the 25 percent statistic accounts for all severe cases."

The tricky thing about a (probably misleading) statistic like that is 0% of the autistic people I know are nonverbal, but I suppose if I know them in any real capacity, they have to be verbal. Anyway, this freakout about autism seems overblown to me. Can any RFK championers explain this one to me?

christianqchung··on GPT-4.1 in the API
Okay, it's common across other industries, but not this one. Here is Google, Facebook, and Anthropic comparing their frontier models to others[1][2][3].

[1] https://blog.google/technology/google-deepmind/gemini-model-...

[2] https://ai.meta.com/blog/llama-4-multimodal-intelligence/

[3] https://www.anthropic.com/claude/sonnet

christianqchung··on GPT-4.1 in the API
But Llama 4 Scout does badly on long context benchmarks despite claiming 10M. It scores 1 slot above Llama 3.1 8B in this one[1].

[1] https://github.com/adobe-research/NoLiMa

← PreviousPage 2 of 5Next →