I love using AI to set up projects in 5 minutes but I hate to develop these projects using AI because inevitably it runs into a wall and I need to guide it and fix its code.
I suppose in this case it picked up an existing project and DIDN’T walk off a cliff? Were the mutations really small?
I’ve been putting questions into LLM research functions, including Claude’s research mode, and letting them churn until a report appears.
I’ve been starting with topics where I’m already familiar with the answer but want a refreshed. So far, I’m not impressed. Some times the info will be correct. Most of the time it strings together a lot of words from the material it finds but it reads like an undergrad trying to paraphrase the Wikipedia page without understanding the content. Often it will have one bullet point that is completely wrong.
The other problem I’m having is that it’s not very good at identifying poor sources. This is less of a problem with topics like math and engineering, but a big problem with topics like health and medicine where it will pick up alternative medicine and pseudoscience pages and integrate them into the research as if they were real. There are a lot of health and medicine topics where the way pseudoscience people talk about a subject doesn’t match the real science, but they use the same words and therefore catch the same search terms.
An example is the way “dopamine” is used in casual conversation and by influencers in ways that aren’t accurate. Concepts like “dopamine fasting” or claiming things “raise your dopamine” aren’t scientifically accurate but use the same words nevertheless and therefore can get pulled into the training set and searches.
Ah, another weekly AI booster article from the AI booster gang.
Hate to break it to you, but if GPT-5 is better AI researcher than you, you were probably not that good to begin with.
Does Codex tell you why 95% of AI projects in the enterprise fail? Or why the only study up to date on merits of AI for coding shows 19% decrease of productivity.
The author wasn't doing "AI research" before and neither was GPT5. This is not at the frontier of anything, it is just an already solved problem in training that GPT5 found. Had the author been willing to actually do a Google and GitHub search, or just twiddle the training knob parameters enough on their own, they would have found a better solution than working alone.
Also this footnote:
> Alone” here is relative - I did use ChatGPT and a bit of Copilot to generate some of the training code in my last attempt. I just didn’t use any agentic tool
I have no words. I wonder if this "AI researcher" can make it through the original Attention Is All You Need paper without an LLM.
So why is "distilling from N-gram" better, why does it make the transformer learn English faster?
Hypothesis: it's the standard "teacher-student" or "distillation" trick - if you're learning next-token-prediction, you only learn what the correct answer is (i.e. the spike in probability), but when you're distilling from a teacher model, you learn the entire distribution of potential answers.
Curious, can anyone more experienced in AI research comment on this?
> OpenAI has released GPT-5-codex, and supposedly uses it ... to automate a lot of their ... AI research
If I was the owner of an AI company that was forever trying to juice its valuation and raise money, you can bet I'd be telling people I had built a magic self-improving AI.
I’ve had the exact same experience. I’ve been vibe coding most of my research now, previously was an MLE handcrafting model code.
A lot of negative comments on here, which seems to always be the case with HN and vibe coding. The reality is that it’s actually starting to work, quite well.
Right now AI is lifting up the floor, so if you don't know programming, mandarin or any other topic it will sure do better than you. (Vibe coding goes here)
Same for tasks you know how to do but AI does them faster, there is also value there. (Claude Code used by a senior goes here)
The interesting thing is when AI is lifting up the ceiling everywhere, but maybe then is when we are almost on AGI territory.
This is written by someone who's not an AI researcher, working with tiny models on toy datasets. It's at the level of a motivated undergraduate student in their first NLP course, but not much more.
If one can easily reach parity with a motivated undergrad by leveraging LLMs I will still consider it impressive.
While the 5-minutes model will never be useful in itself it lays the groundwork for amateurs and small groups to getting into developing small models. There's at the moment another HN headline hyping up a tiny model that scores impressively at the arc-agi benchmarks so it's clearly not a dead end to explore what is "household-affordable" models.
Though an approach that doesn't lean on the authors $200/month OAI sub would've been more interesting to follow.
15 comments
[ 0.21 ms ] story [ 38.3 ms ] threadThey aren't excited about anything. They aren't in awe. They haven't done any hard work. They're just here to ooze lukewarm sludge
I suppose in this case it picked up an existing project and DIDN’T walk off a cliff? Were the mutations really small?
I’ve been starting with topics where I’m already familiar with the answer but want a refreshed. So far, I’m not impressed. Some times the info will be correct. Most of the time it strings together a lot of words from the material it finds but it reads like an undergrad trying to paraphrase the Wikipedia page without understanding the content. Often it will have one bullet point that is completely wrong.
The other problem I’m having is that it’s not very good at identifying poor sources. This is less of a problem with topics like math and engineering, but a big problem with topics like health and medicine where it will pick up alternative medicine and pseudoscience pages and integrate them into the research as if they were real. There are a lot of health and medicine topics where the way pseudoscience people talk about a subject doesn’t match the real science, but they use the same words and therefore catch the same search terms.
An example is the way “dopamine” is used in casual conversation and by influencers in ways that aren’t accurate. Concepts like “dopamine fasting” or claiming things “raise your dopamine” aren’t scientifically accurate but use the same words nevertheless and therefore can get pulled into the training set and searches.
Hate to break it to you, but if GPT-5 is better AI researcher than you, you were probably not that good to begin with.
Does Codex tell you why 95% of AI projects in the enterprise fail? Or why the only study up to date on merits of AI for coding shows 19% decrease of productivity.
Also this footnote:
> Alone” here is relative - I did use ChatGPT and a bit of Copilot to generate some of the training code in my last attempt. I just didn’t use any agentic tool
I have no words. I wonder if this "AI researcher" can make it through the original Attention Is All You Need paper without an LLM.
Hypothesis: it's the standard "teacher-student" or "distillation" trick - if you're learning next-token-prediction, you only learn what the correct answer is (i.e. the spike in probability), but when you're distilling from a teacher model, you learn the entire distribution of potential answers.
Curious, can anyone more experienced in AI research comment on this?
If I was the owner of an AI company that was forever trying to juice its valuation and raise money, you can bet I'd be telling people I had built a magic self-improving AI.
A lot of negative comments on here, which seems to always be the case with HN and vibe coding. The reality is that it’s actually starting to work, quite well.
Same for tasks you know how to do but AI does them faster, there is also value there. (Claude Code used by a senior goes here)
The interesting thing is when AI is lifting up the ceiling everywhere, but maybe then is when we are almost on AGI territory.
We have several synthetic datasets and automated evaluation options for such things that were close to impossible to do before LLMs.
While the 5-minutes model will never be useful in itself it lays the groundwork for amateurs and small groups to getting into developing small models. There's at the moment another HN headline hyping up a tiny model that scores impressively at the arc-agi benchmarks so it's clearly not a dead end to explore what is "household-affordable" models.
Though an approach that doesn't lean on the authors $200/month OAI sub would've been more interesting to follow.