It is believed dense models cram many features into shared weights, making circuits hard to interpret.
Sparsity reduces that pressure by giving features more isolated space, so individual neurons are more likely to represent a single, interpretable concept.
Yes, although the sparsity doesn't need to be inherent to the model - another approach is to try to decode the learned weights using approaches like sparse auto-encoders or transcoders.
I'm also very excited about SAE/Transcoder based approaches! I think the big tradeoff is that our approach (circuit sparsity) is aiming for a full complete understanding at any cost, whereas Anthropic's Attribution Graph approach is more immediately applicable to frontier models, but gives handwavier circuits. It turns out "any cost" is really quite a lot of cost - we think this cost can be reduced a lot with further research, but it means our main results are on very small models, and the path to applying any of this to frontier models involves a lot more research risk. So if accepting a bit of handwaviness lets us immediately do useful things on frontier models, this seems like a worthwhile direction to explore.
The OpenAI fine-tuning api is pretty good - you need to label an evaluation benchmark anyway to systematically iterate on prompts and context, and it’s often creates good results if you give it a 50-100 examples, either beating frontier models or allowing a far cheaper and faster model to catch up.
It requires no local gpus, just creating a json and posting to OpenAI
One interesting anecdote about this bill was that the European Commission allegedly funded digital advertisements promoting it, targeting specific political demographics, which is something that could possibly be prohibited by their own regulations.
HHI is a pretty interesting metric. It’s calculated by taking each firm’s market share, squaring it, and summing across all firms.
This gives the probability that two randomly chosen customers belong to the same firm.
In one micro models of oligopoly, Cournot competition, it lines up directly with the markup firms can sustain.
Outside of theory, it’s an intuitive way to average together the market power of all firms, with increases in market share for bigger players being weighted more heavily.
Pretty interesting - some contamination, some better answers, and it failed to write a sentence with all 5-letter-words. I’d have expected it to pass this one!
There is a big upside potential for high growth companies taking advantage of technology trends.
Today, Google’s revenue is £263.66 Billion. This is nearly 300x the revenue Google generated in 2003 ($961.9 million). The company went public on August 19, 2004, at $85 per share, valuing the company at $23 billion. After the IPO, Google reported $1.47 billion in revenue for fiscal year 2003, with a profit of $105.6 million.
But let’s ask a different question: aside from re-allocating the economy’s marketing and advertising budget into Google (from, presumably, local newspapers and TV before Google existed) how much of that revenue comes from actual tangible new wealth creation?
To put some context on this, 78% of Google’s revenue is advertising. Overall US ad spending has been increasing at about 1.6% per year since 2001, with no obvious indication of an acceleration (beyond some bumps around 2007/8.) So is there actually a success story beyond market capture here? And if all we’re doing is concentrating existing business into new channels, is this something we should be excited about?
You simply pay more for a product to find you. The more overpriced ones find you first. Googles business model is to make it as hard as possible for products to find you while simultaneously pretending to be the go-to place for precisely the opposite. A truly magical accomplishment.
Google created android - the most popular OS. Sure maybe samsung or nokia would be used instead but definitely the helped expend ad business with android. Same like Meta/ByteDance expanded Ad business with Intagram/Tiktok. Even if ad spending grew 1.6% per year it's not sure if it grew as much if android didn't exist. Also need to take into account probably reduced cost of advertising - this product just got cheaper. That the ad market grew 50% in 25 years doesn't mean we have only 50% ads served same like 50% grow (in $) in smartphone market doesn't have to mean you have only 50% more smartphones if they got cheaper.
Technically, Andy Rubin and Chris White weren't at Google when creating Android. In usual big tech fashion, Google did a good acquisition but didn't actually "create" Android, they bought it.
This is an interesting topic and I’m not sure it has an answer. In 1995 advertising was really spray and pray. Testing ads was a really difficult proposition so we saw things like bearer coupons (mention this ad or bring in this coupon for 20% off!). The dominant advice out of radio was to play an ad constantly. That advice worked well for traditional media but not so well for advertisers. That model worked so well for advertisers that thirty years later, people around my age in my city can all sing the same five advertising jingles.
Google provided a toolkit to test ads and figure out which are most effective. Now the other side of that argument is that in industry, a massive of percentage of qualified people still spray and pray. The advertising industry as a whole is far from data driven.
At one point, there was an argument this was good for the planet. My newspapers are much thinner than they were 30 years ago when I could collect a metre of newsprint a month if I subscribed to the Globe and Mail plus a local. But I don’t think anyone can claim now that data centres are environmental miracles. This has also decimated local journalism to such a point that people are less aware of environmental catastrophes in their own relative backyards.
It’s possible the net effect was positive and advertising is more efficient. It’s more accurate to say advertisers have a toolkit to analyze effectiveness but many don’t or aren’t capable.
Edit - I’m going to give a very specific example of a radio jingle. If anyone is around forty or older and from a major city in Saskatchewan, they will be able to finish this.
“I said no no no no don’t pay anymore, no GST and no money down.”
Maybe another way to say this is: do targeted ads materially advance society? Is there an argument that better ad-targeting has increased GDP, or improved overall economic growth in some other way? Would a less-efficient online advertising system produce dramatically worse outcomes? And did we get more or less back from the older system (TV+local journalism) than we get from Google?
But is a Google with 78% market share actually important to this? Could we have a network of companies doing this job, perhaps less efficiently, and the world would be just about as decent?
Is that because of innovation? Or is that because of Google’s antitrust activity the US government is currently busting? Safari default search engine deal, etc.
"Antitrust" is just lawyer-talk for winning strategies that we later arbitrarily decide is not good for capitalism.
They weren't a bunch of gremlins in a cave conspiring to commit "anti-trust violations" in 2005. They were smart as hell and invested in the right areas.
Microsoft would get hit with the same anti-trust Google is being hit with if Bing and Windows Phone were successful - they're getting away with it because they're terrible.
Anti Trust is not lawyer speak for winning strategies. It’s a specific term, and theres a time and place to use it. Anti trust is what people, especially programmers, have been saying from time immemorial when these firms became this big.
Inefficient markets are bad for humans and are bad markets. The allocate resources inefficiently. The google graveyard is (arguably) a case in point.
The reason Khan reached the FTC was because her thesis at college, made the case that amazon’s actions reduced customer welfare. A fact that was covered here, on HN. This isn’t something a community notices unless it matters to them.
> They weren't a bunch of gremlins in a cave conspiring to commit "anti-trust violations" in 2005. They were smart as hell and invested in the right areas.
You seem to possess the (very common) misconception that monopolies are illegal. They are not. Rather, it is illegal to intentionally use one's monopolistic position to make it difficult or impossible for others to compete.
Winning for who? Not for society as a whole, that's certain.
To put it in money terms so even you can understand it, how much time has been wasted globally because Google is peddling ads and spam sites instead of pointing people to useful results?
Is that free? We should substract it from the GDP calculations if you ask me...
The meaning of “antitrust” is very clearly defined. You’re allowed not to like it or to think that laws against it should be struck down, or whatever, but you can’t say it’s something it’s not.
Certainly, when a law that wasn’t applied begins to be applied again, I can see certain mental models taking issue with that. Regardless, the law and its historical application and consequences didn’t change, only a politically driven low enforcement period. That was the anomaly.
If you allow one company to achieve market dominance, it suffocates the ecosystem and stifles evolutionary growth pressures. It's concentrated malinvestment into a local maxima that salts the playing field so thoroughly that escape velocity is unattainable by anyone else.
There are models of this. And historical anecdotes and evidence.
Google search, google maps, gmail, YouTube, and Chrome have all been good functional products for over a decade. I genuinely don’t know what they’ve been doing since then other than milking us and getting new customers. Maybe 10% of this growth leads to a real improvement in human lives.
A. below is a list of OpenAI initial hires from Google. It's implausible to me that there wasn't quite significant transfer of Google IP
B. google published extensively, including the famous 'attention is all you need' paper, but open-ai despite its name, has not explained the breakthroughs that enabled O1. It has also switched from a charity to a for-profit company.
C. Now this company, with a group of smart, unknown machine learning engineers, presumably paid fractions of what OpenAI are published, has created a model far cheaper, and openly published the weights, many methodological insights, which will be used by OpenAI.
1. Ilya Sutskever – One of OpenAI’s co-founders and its former Chief Scientist. He previously worked at Google Brain, where he contributed to the development of deep learning models, including TensorFlow.
2. Jakub Pachocki – Formerly OpenAI’s Director of Research, he played a major role in the development of GPT-4. He had a background in AI research that overlapped with Google’s fields of interest.
3. John Schulman – Co-founder of OpenAI, he worked on reinforcement learning and helped develop Proximal Policy Optimization (PPO), a method used in training AI models. While not a direct Google hire, his work aligned with DeepMind’s research areas.
4. Jeffrey Wu – One of the key researchers involved in fine-tuning OpenAI’s models. He worked on reinforcement learning techniques similar to those developed at DeepMind.
5. Girish Sastry – Previously involved in OpenAI’s safety and alignment work, he had research experience that overlapped with Google’s AI safety initiatives.
> A. below is a list of OpenAI initial hires from Google. It's implausible to me that there wasn't quite significant transfer of Google IP
I agree there's hypocrisy but in terms of making a strong argument, you can safely remove your list of persons who (drum roll)... mostly _didn't_ actually work at Google?
I think this project is awesome and am quite disappointed with some cynical commentary from large American labs.
Researcher at Meta or OpenAI spending hundreds of millions on compute, and being paid millions themselves, whilst not publishing any of their learnings openly, here a bunch of very smart, young Chinese researchers have had some great ideas, proved they work, and published details that allow everyone else to replicate.
"No “inscrutable wizards” here—just fresh graduates from top universities, PhD candidates (even fourth- or fifth-year interns), and young talents with a few years of experience."
"If someone has an idea, they can tap into our training clusters anytime without approval. Additionally, since we don’t have rigid hierarchical structures or departmental barriers, people can collaborate freely as long as there’s mutual interest."
Lendable | Data Science & Software Engineering |Hybrid & Remote | UK, London | Full Time
Lendable have built the big three consumer finance products from scratch: loans, credit cards and car finance. Help us improve the best-in-class credit analytics that powers credit-decisioning.
This is a really nice open source coding agent implementation. The use of async is interesting.
reply