Ask HN: Which sources cover AI developments without falling for the hype? | Hacker News Reader

Ask HN: Which sources cover AI developments without falling for the hype?

126 pointspramodbiligiri82 comments

Popular press seems to do a bad job covering AI related developments. Twitter is too scattered and academic papers are too narrow.

So where do working professionals get seasoned and mature coverage of this space? What would be the AI equivalent of the Economist, AnandTech or Tom’s Hardware?

View on HN

I've been trying to do this on my blog. Here's my AI tag: https://simonwillison.net/tags/ai/

I really like your blog!

FWIW, and to answer the OP's question, I think yours and Ethan Mollick's blogs and social media accounts are the best sources for AI news right now. I've been trying to find more sources myself and, other than a few good but highly technical ones that don't focus on actual LLM use (e.g. Jeremy Howard), have bern coming up short.

You, Ethan, & swyx are fantastic.

thank you so much! indeed trying to keep a level head on while the world goes nuts, and consciously missing out on the social media dopamine by not endorsing every little thing as lifechanging. gotta say the creator incentives arent set up to encourage it, I think all 3 of us purely do it for love of the game rather than making any serious money from our work on this

Thanks for this! IMHO it's a bit wall-of-text (or maybe that's the weakened attention span). Ever considered present just headlines like a certain tech news site?

Yeah, that's a very reasonable improvement I should make. My homepage is modeled after https://daringfireball.net/ and has been for the past 20-ish years, but it's time I redesigned that to better reflect my content I think.

Seems fine via feed reader but in my case, DF is pretty much the only thing I consistently read via feeds since Google Reader was axed... :-)

Please don't make that "improvement". This presentation is, and signals, real content instead of content mill.

Perhaps feature a selected tag nav above your content, for ability to jump into a topic net faster, more scannable with better discrimination of content to dive into versus continue scanning past. That shows the depth of content, and someone can bookmark a tag or set of tags to get what the post above is asking for.

Keep the DF style home page content for those who have a bookmark to your site and read what's new, or an RSS feed to your site. (Though RSS can work different from home page, of course, and should remain long form regardless.)

The recent progress in “AI” is almost entirely due to advancements in large language models (LLMs).

In some aspects the hype is real; LLMs are extraordinarily performant for a wide range of previously hard tasks.

On the other hand, people seem to equate these advancements with “strong AI” (or AGI). We are one step closer, sure, but the calculator was also a step forward.

We’ve created a mirror of all (most) human knowledge, queryable via natural language. People look into this mirror and see themselves, sometimes things greater than themselves.

This mirror tricks us into thinking the machine will soon replace us. It’s so accurate, why would it not?

Fortunately, it’s just a mirror, and we’re the bear in the woods seeing it’s reflection for the first time. Scared and ready to fight.

If you focus on the technology (LLMs) and throw caution at anyone hyping “AI” generally, you can create a filter for what’s real and what should be questioned.

There are research papers left and right.

Stuff like certain architectures, reading LLM for promoting multi modal llms.

Then we have stuff like insteuctgpt, ml models for robots, lots and lots of research from Nvidia for virtual simulation and transfer to real world, digital twin is also a relevant art in agi.

Object detection is also much better and has nothing to do with llms. Segment anything from FB for example.

Whisper and sd are also not LLM.

There are a ton of puzzle peaces slowly falling in place left and right.

"There are a ton of puzzle peaces slowly falling in place left and right."

Yet, we do not seem to have a very good understanding of how many pieces there are in the puzzle.

True.

But I feel well entertained watching them fall. Like using them and experimenting around.

But it also shows the road ahead quite clear. For example were is the money coming from? From millions of people paying for GitHub copilot for example.

How is it sold? Per webui, API and cloud providers.

Digital twin will also play a huge role in this as a bridge between AGI <> real world.

They may not be "large" in the same sense that GPT4 is "large" but apart from then simulator stuff, every single one of the models you mentioned is transformer-based. Every one of them basically includes encoders to project different modes of information (images and audio) into a "language-like" space so that it can be compared with and mapped to and from text. I think it's fair to say that language models, if not LLMs, unlocked a surprising amount of power.

The issue here is "It's complicated"

For example, looking at the mechanical replacement of human strength in the 1800 and 1900s shows people that the human hardship costs where real. The labor wars in the US are a good example of this. The process of mechanization shifted power to the hands of the capitalists, and was only wrestled back with blood.

The real key of the future with AI will be the question of generalization. Multimodal AI does show a reasonable amount of ability on predicting real world events. For example, show a picture of a kid opening a bike and ask what is next in image form, and the AI will return a picture of the kid riding a bike. This ability of reasonable prediction based on sets of 'real world' input is not something that we've had in previous generations of computer systems. Again, if these systems generalize well, rapidly become cheaper, and enable the capitalist class to gain more wealth expect their use to explode at a near exponential rate.

Very few reasonably educated people say "AI will never reach human ability", the only question that is really being asked is when, and in a lot of peoples eyes when has moved much sooner.

definite "maybe" - a model can only return elements present in the training material. This is a powerful formula, but not "everything" .. blurring the story with a child, and prediction as a general quality.. is moving into "deception technique" either consciously or not IMHO

> We are one step closer, sure, but the calculator was also a step forward.

Even that isn’t particularly clear, I don’t think. A speculative future AGI probably won’t be a fancy LLM, or at least there’s no particular reason to think it would be.

People use "AGI" in different ways. The term has vague meaning. Some mean true intelligence. Some mean it can wash their dishes.

That said, technology generally improves exponentially. So, where will we be in 5 years?

I follow these on YouTube:

* Matt Wolfe: https://www.youtube.com/@mreflow

* MattVidPro AI: https://www.youtube.com/@MattVidPro

* Two Minute Papers: https://www.youtube.com/@TwoMinutePapers

* Dr Alan D. Thompson: https://www.youtube.com/@DrAlanDThompson

* Curious Refuge: https://www.youtube.com/@curiousrefuge

I can't speak for the others but Two Minute Papers' general format oozes hype. I realize that he tends to recap on exciting topics, but he also sometimes "falls for" research that won't actually pan out for practical reasons. He also tends to cover things in an intensely over-optimistic way ("WOW! WHAT A TIME TO BE ALIVE" on literally every video...) that would have you think every single subject is an absolute major breakthrough when the reality is that many of them simply produced cool-looking demos that look good in a youtube thumbnail.

Sorry if that's too harsh. The channel isn't all bad. The title of the channel is up front about being a quick synopsis that you can watch from time to time to keep up with the latest updates. That's still useful for many people. But it probably won't work as a good _filter_ for what will and will not be quickly forgotten.

In fairness (again), that's a _very_ difficult challenge to solve. It is however the premise of the Ask HN.

I watch him for a long time and he is just right.

The progress on so many fields is fast.

And a ton is already available to use.

All the nerfs, most of Nvidia topics, etc.

Or all the character animation stuff, those things are in games and in the industry.

And even the few which show the direction make it very obvious were the road is going in the next 5-10 years.

I'm honestly surprised that you think he is naive/to hyped.

Alone the Nvidia ml denoiser and real time ray tracing is basically redefining graphs. We are in the middle of all of it.

Check out this channel if you've not seen it before

https://www.youtube.com/@aiexplained-official

Also robert miles, though most of his videos are older at this time

https://www.youtube.com/@RobertMilesAI

Definitely Two minute papers. I'm not a huge fan of Matt wolf because he shills pretty hard for Leonard AI (a stable diffusion SaaS) without being upfront about the fact that he's very likely commissioned by them. It makes me call into question his objectivity.

Substack ?

AI Supremacy: https://bit.ly/3Qz8uNV

Latent Space: https://bit.ly/469AAFd

Encyclopedia Autonomica: https://bit.ly/3FXlVlU

Deep Learning Focus: https://bit.ly/40Bi5bF

Artificial Fintelligence: https://bit.ly/3SxQNAZ

Waiting for the obligatory saint, who answers with the actual URLs.

AI Supremacy: https://aisupremacy.substack.com/

Latent Space: https://www.latent.space/

Encyclopedia Autonomica: https://jdsemrau.substack.com/

Deep Learning Focus: https://cameronrwolfe.substack.com/

Artifical Fintelligence: https://finbarrtimbers.substack.com/

Unfortunately, Latent Space is not RSS friendly

The RSS feed for it here appears to work: https://www.latent.space/feed

latent space author here - its a substack, it should work

let me know if any other issues!

Whats the point of the bit.ly links?

i'm guessing https://support.bitly.com/hc/en-us/categories/360001889231-A...

https://aisupremacy.substack.com/ https://www.latent.space/ https://jdsemrau.substack.com/ https://cameronrwolfe.substack.com/ https://finbarrtimbers.substack.com/

AI Supremacy: https://aisupremacy.substack.com/ Latent Space: https://www.latent.space/ Encyclopedia Autonomica: https://jdsemrau.substack.com/ Deep Learning Focus: https://cameronrwolfe.substack.com/ Artificial Fintelligence: https://finbarrtimbers.substack.com/

and I did not click on the links, but used a search engine to find the sources.

The latent space podcast is really good. They have some top notch guests and ask insightful questions. They also feel pretty grounded compared to a lot of the hype.

Others already posted the resolved links, but out of interest I tried resolving them using ChatGPT which worked surprisingly well:

https://chat.openai.com/share/29c10bb1-9576-43aa-9a19-e672f3...

I had the same problem, so I ended up classifying all HN posts because I believe HN is my most relevant and trustworthy source for tech news. Example: https://www.kadoa.com/hacksnack/6194542d-2157-4e3c-8321-a437...

This was more of an experiment for a personalizable HN feed, but I'll fully productize it if there is enough interest.

Hey, this is really interesting. Out of curiosity -- I'm assuming you scrape the front page regularly, classify the posts, then sort them accordingly and do the summary and comment report via LLMs, right?

In any case, I'm definitely interested in this and I can see myself using it fairly often.

Yeah that's pretty much how it works :) will launch this as Show HN soon.

This is neat, the comment summary metrics are a nice touch. I'd use it.

Oh man, I wish this thing had an RSS feed... would be perfect.

Will add RSS, good idea!

I find Zvi Mowshowitz to be the best source available (https://thezvi.substack.com/). He’s in the X-risk camp (and so am I) but seems to have a clear view of some AI things being exciting, others being dangerous, and still others being irrelevant.

I don’t trust most AI-positive sources because they almost never have anything negative to say at all, so they’re clearly in it to hype AI and not to inform anyone of true things. I don’t trust Gary Marcus’s opinion for a similar reason.

X-risk camp?

Existential risk, I think.

https://www.lesswrong.com/tag/existential-risk

Ah so like a Yudkowsky type then? I don't want to lump them all into the same category, but since "X-risk" is given as a category I would just say that they're likely all very well educated, and for some reason deeply deeply misguided by some sort of philosophical hunch that is not rooted in actual current research but in thought experiments. Yudkowsky tends not to respond well or at all to criticism either. For real just go see the criticisms written on that site about his views on runaway AGI and how we get "one critical chance" to stop it from killing us all. His response is effectively "this post is very long, could you summarize the best bits?" and then doesn't respond any further.

Less on topic, I'm becoming convinced that lesswrong is where researchers who couldn't get published go to feel like they got published and peer reviewed, when in reality they are getting peer reviewed by a massive echo chamber. Even less on topic and potentially downvote-inciting; it looks less like an echo chamber and more like a sort of social-media-first cult every day.

By X-risk person I just mean someone who thinks AI does have near-existential risks, and those risks are significant enough to merit being a top consideration in making decisions about AI.

I don’t personally base my stances on Yudkowsky’s ideas, and never got into Lesswrong.

Zvi is definitely not like Yud, he's much more balanced. He is concerned about X-Risk but he seems to give reasonable space to the opposing arguments. He has a specific section "Other People Are Not As Worried About AI Killing Everyone" for this in his roundups.

His roundups also cover a lot of things going on in AI and don't just hyperfocus on safety as the one and only important thing.

Hinton and Bengio are also in the x-risk camp, if you want an appeal to authority.

humans using AI in competing teams will cause the destruction of ${CIV}

Following AI-related RSS feeds from the ML research arms of the big companies is nice. Not as hypey as the popular press, and you get more technical details. Most mail clients support RSS nowadays, so you can even get notified when a new post is made, instead of periodically checking on Mondays or something. IBM, Microsoft, Meta (I don't think they have RSS though), nVidia, OpenAI, Google, and IBM all have great blogs that cover their work.

Another good resource is the YouTube channel "2-minute papers." It sometimes has a lot of hype, but it does a good job of showcasing recent work.

Surprised no one has mentioned AK’s daily paper digest: https://huggingface.co/papers

He also posts summaries on Twitter, or at least he used to but my Twitter account is glitched and I can’t see Tweets anymore

In general, any new research from OpenAI that is subsequently quickly replicated by the actual open source community (e.g. Latent Consistency Models for a recent example). They have put out research that hasn't gone anywhere in the past - which is why I suggest waiting until the open source community or other AI startups start to copy them on that specific research.

Another group that is important to watch is any of the members from the `CompVis` group that originally developed VQGAN and Latent Diffusion models. Although I'm uncertain how much of the team remains as many seem to have realized they can do more research (and make some more money) by working at the various research labs popping up.

Those outlets cover mature industries. And as Charles Kettering tells us (inventor of the electric starter motor): "You can’t plan industries". Of course no one is yet out in front with the settled science to preach so just enjoy the chaos and behind the scenes look at the birth of an industry or use an aggregator.

https://allainews.com

https://nuse.ai

https://news.bensbites.co/newest

https://trendingpapers.com/

Pure numbers: the top trending papers surface. They are a function of PageRank (citations and the importance of which papers cite each other), authors' previous body of work, etc...

The filters help select a sub-area (NLP, Computer Vision, etc.) and slice what's really new (released over the last week, last 3 months, last 6, etc.).

The tool is designed to solve this problem.

That's a good one, and in a similar vein there's https://paperswithcode.com/ from facebook/meta that tracks papers and their accompanying github repos, along with their benchmark results.

Twitter is good but you have to filter it. Start by following AI researchers doing real work, not pundits. However that's not enough by itself. It is essential to use the "mute words" feature to aggressively prune any and all tweets about politics and other crap. You're left with a much denser stream of people's informed thoughts about the latest research as it comes out.

How is this not a default in the featuerset?

You’re still in the mindset that Twitter is there to help you keep up with people and topics you care about, instead of… whatever the hell Elon is trying to do with it this month.

I was looking for something similar as well since the initial SD launch. But I've ended up just using the HN frontpage as that, with good results. Most things end up here, even if not driven by hype, and the discussions that follow usually are "Good Enough" but with quite a bit of off-topic comments.

Zvi does a very detailed roundup of the weekly news, with an AI x-risk lens:

https://thezvi.substack.com/

Pretty thorough (though verbose) if you just want to stay on top of developments.

On the AI policy side: https://aipoliticalpulse.substack.com/

AI ethics stuff mostly: https://www.machine-ethics.net

https://thegradient.pub/

is not bad.

The term "artificial intelligence" is intrinsically hype.

That's why it is always a moving bar.

Good luck.

I think that when people accept that a machine can be intelligent, the goalpost will stop moving, but of course it will take time.

> when people accept that a machine can be intelligent, the goalpost will stop moving

The goalposts will stop moving when the terms used are precise. What does intelligence mean? If people agree on a precise definition then we’ll know when we’ve reached that artificially.

That would be difficult because intelligence is not black and white, so an arbitrary line should be decided, but at least everyone agrees that humans are intelligent. Maybe people will finally stop moving the goalpost when AI would be able to beat humans on every tasks.

well, given that industry people and “experts” confidently predicted that what has happened in 2023 could happen only hundreds of years from now or not at all… i would say that the “hype” has as much veracity as anything else. the simple fact os that no matter how carefully you curate your news, you will have no idea whats coming. and coming soon. most people havent wrapped their head around it yet but the plain and obvious truth is that technology, especially AI, must be slowed, paused, regulated in some way because too much change too fast is very dangerous.

please dont comment about previous eras of technology and change. nothings even comes close to comparing

High level AI developments:

https://importai.substack.com/

https://www.semianalysis.com/ Hardware focused. Paywalled, but long article teasers contain plenty of information.

Research Focused:

https://codingwithintelligence.com/

https://dblalock.substack.com/

https://twitter.com/arankomatsuzaki

Reddit local llama

Andrej Karpathy

Simon Willison

Ethan Mollick

Riley Goodside

Matthew Berman (for succinct howto's on YT)

https://simonwillison.net/

At this point it's better to look for specific people rather than a single source. Some suggestions

Gary Marcus https://garymarcus.substack.com/

Temnit Gebru and Dr. Emily Bender https://www.dair-institute.org/

Alex Hanna, Mystery AI Hype Theater: https://www.buzzsprout.com/2126417

Dr. Émile P. Torres: https://www.xriskology.com/

Gary Marcus covers Gary Marcus

lol all the sources you linked have consistently been wrong at prredictions over the last 2 years

If they are constantly wrong they are a good source. Just take the opposite

The opposite of a wrong prediction is not necessarily true. It depends on the form of the statement.

What makes Gary Marcus a "Leading expert on AI" (in his own words)?

These sources all fall for a very different kind of AI hype.

Hilarious answer

Missing Elon Musk, guy is a leading authority on all things AI that even the British government asks what to do about it from him (big /s).

And the podcast guy that speaks really slowly ...

Reply on news.ycombinator.com