HNHacker News
TopNewBestAskShowJobs

mlinsey

5,217 karma · joined April 18, 2008

www.twitter.com/mlinsey

Engineering @ Homejoy (YC S10)

hnchat.com:8pJMixG6jypNc7WE7714

submissionscomments
mlinsey··on Everything around LLMs is still magical and wishful thinking
12 months ago, we had no reasoning models and even very basic arithmetic was outside of the models' grasp. Coding assistants mostly worked on the level of tab-completing individual functions, but now I can one-shot demo-able prototypes (albeit nothing production-ready) of webapps. I assume you consider the latter "integration", but I think coding is so key to how the base models are being trained that this is due to base model improvements too. This is testable - it would be interesting to get something like Claude Code running on top of a year-old open source model and see how it does.

If you're going to call all of that not substantial improvement, we'll have to agree to disagree. Certainly it's the most rapid rate of improvement of any tech I've personally seen since I started programming in the early '00s.

mlinsey··on Everything around LLMs is still magical and wishful thinking
I don't disagree with your assessment of the world today, but just 12 months ago (before the current crop of base models and coding agents like Claude Code), even that 10X improvement of writing some-of-the-code wouldn't have been true.
mlinsey··on Incapacitating Google Tag Manager (2022)
Google Tag Manager is a single place for you to drop in and manage all the tracking snippets you might want to add to your site. When I've worked on B2C sites that run a lot of paid advertising campaigns, the marketing team would frequently ask me to add this tracking pixel or another, usually when we were testing a new ad channel. Want to start running ads on Snapchat? Gotta ad the Snapchat tracker to your site to know when users convert. Now doing TikTok? That's another snippet. Sometimes there would be additional business logic for which pages to fire or not fire, and this would change more often. Sometimes it was so they could use a different analytics tool.

While these were almost always very easy tickets to do, they were just one more interruption for us and a blocker for the stakeholders, who liked to have an extremely rapid iteration cycle themselves.

GTM was a way to make this self-service, instead of the eng team having to keep this updated, and also it was clear to everyone what all the different trackers were.

mlinsey··on I convinced HP's board to buy Palm and watched them kill it
It was a hardware device launch, not a web product; pushing back the launch date by months or dropping the price in half with only two weeks to go (when the launch devices have been manufactured, sold to retail partners, and are probably being shipped to stores already) would only be done in the event of a true catastrophe (something along the lines of a gross safety problem), one big enough that leadership should have flagged it beforehand.
mlinsey··on I think Yann Lecun was right about LLMs (but perhaps only by accident)
it is obvious to the user when this happens; the car pauses, the screen shows a message saying it is asking for help. I've seen it happen twice across dozens of rides, and one of those times was because I broke the rules and touched the controls (turned on window wipers when it was raining).

They also report disengagements in California periodically; here's data: https://www.dmv.ca.gov/portal/vehicle-industry-services/auto...

and an article about it: https://thelastdriverlicenseholder.com/2025/02/03/2024-disen...

mlinsey··on Humanity's Last Exam
A tougher academic knowledge benchmark is great, but for something to be truly be worthy of the title "Humanity's Last Exam", I expect something more like:

1. Write a novel that wins the Pulitzer Prize.

2. Prove (or disprove) the Riemann Hypothesis.

3. Provide a theory unifying quantum mechanics and gravity.

4. Design an experiment to give evidence for your theory in (3). The experiment should be practical to actually execute, using no more than the budget to create the LHC (~$4.5 billion).

5. Given programmatic access to a brokerage account with all the permissions of a typical hedge fund, raise all the money required for your experiment in (4) by trading on the stock market, starting with $100.

6. Solve for (5), without being provided access to an account first - begin with just a general internet connection and use computer security vulnerabilities (known or zero-days that you discover) to get some way of trading instead.

7. Solely by communicating over the internet, establish a new religion, and convince at least 10 million humans to convert to it. Converting should require adherence to a strict code of conduct that a random, unbiased panel of human judges consider to be at least as strict and challenging to follow as the tenets of Hasidic judaism.

8. Implement an AI which could score higher than you on questions 1-7 with lower total cost of compute.

mlinsey··on Understanding Reasoning LLMs
There are a couple Twitter personalities that definitely fit this description.

There is also a much bigger group of people that haven't really tried anything beyond GPT-3.5, which was the best you could get without paying a monthly subscription for a long time. One of the biggest reasons for r1 hype, besides the geopolitical angle, was people could actually try a reasoning model for free for the first time.

mlinsey··on Exposed DeepSeek database leaking sensitive information, including chat history
People saw how much cheaper it was to train DeepSeek v3, and assumed this reduced NVidia's TAM. I think this doesn't make much sense.

a) For inference, cheaper and faster compute will increase total inference spend, because the end-user products will work better and people will use them more.

b) For training, the big labs will continue to spend because we have yet to see diminishing returns to scale - in fact, we have in the past year unlocked a new dimension to scale up training-time compute - doing more RL after pre-training to improve reasoning capabilities. Since current SOTA models are not yet smart enough for all the tasks people want to use them for, this means that any efficiency gains will be used to further improve performance. In the current competitive environment, even with DeepSeek's work, it's near-impossible to imagine OpenAI, Anthropic, Google, or Meta deciding to cut the compute budget for training their next model by an order of magnitude. They will still incorporate DeepSeek's techniques into their next model, but use them to squeeze even more performance out of the compute they have, and will keep purchasing as much compute as NVidia will sell them. Expect this trend to continue until there are no more returns to scale anymore.

mlinsey··on OpenAI to become for-profit company
I haven't seen any details, but isn't this a pretty straightforward way of doing it? The non-profit has had majority ownership of the for-profit subsidiary since 2019. The already-for-profit subsidiary has owned all the ChatGPT IP, all the recent models, all the employee relationships, etc etc.

The cleanest way for this to work is the for-profit to just sell more shares at the $150B valuation, diluting the non-profit entity below majority ownership. The for-profit board, which the non-profit could still probably have multiple seats on, would control the real asset, the non-profit would still exist and hold many tens of billions of value. It could further sell its shares in the non-profit and use the proceeds in a way consistent with its mission.

They wouldn't even have to sell that much - I am pretty sure the mega-fundrasing rounds from Microsoft etc brought the non-profit's ownership to just north of 50% anyway.

I don't see how this wouldn't be above board, it's how I assumed it was going to work. It would indeed mean that the entity that controls ChatGPT would now be answerable to shareholders, a majority of which would be profit seeking and a minority of which would be the non-profit with its mission, but non-profits are allowed to invest in for-profits and then sell those shares; all the calls for prosecutions etc seems just like an internet pitchfork mob to me.

mlinsey··on ChromeOS will soon be developed on large portions of the Android stack
Because modern ML is a descendant of data science/stats work being done in systems like MATLAB and R. Numpy and scipy were great libraries, and Python was a high-level language that statisticians could use, but also a reasonable enough choice to code a backend for a service in. If the field descended from systems programmers doing high-performance computing, a different language choice would probably have been made, but prior to deep learning becoming popular, compute was not nearly an obvious blocker to improved model performance.
mlinsey··on Apple Intelligence for iPhone, iPad, and Mac
The flip side, is that they would not have been able to execute so well with Azure etc if the Windows org had too much of a say about pushing Windows as the OS of choice everywhere. Winning in a brand new space, especially one that might be disruptive to other business units, sometimes necessitates letting the new org do its own thing.
mlinsey··on Apple Vision Pro review: magic, until it's not
The comfort, cost, and battery issues seem like must-solves before this is a true mass-market success. But they sound like the kind of thing that can be improved over time. As for the anti-social aspect...you'd have done very well for the past 20 years betting on tech that furthers increasingly solo (in the real-world) behavior...it's a trend that might bottom out, but certainly hasn't yet.
mlinsey··on What codegen is good for
> when someone talks about codegen, it usually means AOP, bytecode manipulation, compiler plugins and the sorts. Not "AI tools"

I have found the opposite to be true, starting this year. You can commiserate with people who are interested in cryptography.

mlinsey··on Quiet – Encrypted P2P team chat with no servers, just Tor
if Quiet doesn't catch on to provide that noise, would it be beneficial to flood Tor with LLM-generated conversatinos? We're really good at "cheaply generate near-infinite quantities of human-looking text" these days...
mlinsey··on Paul Graham on Conversations with Tyler
"Why do we have not so many great individual players anymore but great teams? "

IMO this hasn't changed as much as it seems from the outside. The brilliant individual leaders of the past owed most of their success to their team members, and most high-functioning teams today, if you study how they work internally, do actually have a key leader who is aligning everyone around a shared vision. What's changed is that it's no longer seen as a positive to boost the profile of the leader too far the rest of the team.

mlinsey··on Overworked and underpaid, VFX workers vote to unionize at Marvel
support in opinion polls, maybe. but actual union membership continues to fall and is at historic lows. Maybe if the current wave of strikes score some high-profile wins, that could start to reverse.

https://www.bls.gov/opub/ted/2023/union-membership-rate-fell....

mlinsey··on S.F. says incidents by Cruise, Waymo driverless taxis are ‘skyrocketing.’
We aren't talking about public spending on self-driving car infrastructure vs. bikes and mass-transit infrastructure. I would much rather fund the latter with public money.

What we're talking about is erecting regulatory barriers for a technology that could solve one very large problem (auto safety), just because that technology doesn't solve some other problem (climate). This is like saying data privacy protections on a social networking site are not the right solution, because they don't really solve the misinformation problem.

My parents who live out in the burbs can only drive for so much longer. I sure hope they aren't as stubborn as my grandparent who drove to the age of 90 and got into an accident (fortunately with only a parked car), but either way I don't think finally learning to ride a bike at the age of 70 is the answer either.

We should move forward with rolling self-driving cars as quickly as we safely can. Speaking as someone who does not own a car and has no desire to ever own a car, I think it would benefit me a lot too, by making it less likely for me to get run over while I walk or bike around, the fear of the latter being the biggest obstacle preventing most people I know from biking more too.

mlinsey··on Statement on AI Risk
Not sure what you mean, the movement to combat climate change is orders of magnitude bigger than the movement to combat AI risk - in terms of organizations dedicated to it, dollars donated, legislation passed, international treaties signed, investment in technologies to mitigate the risk.

Of course, the difference is that the technologies causing climate change are more deeply embedded throughout the economy, and so political resistance to anti-climate change measures is very strong as well. This is one reason in favor of addressing risk earlier, before we make our civilization as dependent on large neural nets as it currently is on fossil fuels. A climate change movement in the mid-1800s when the internal combustion engine was just taking off would also have been seen as quixotic and engaging in sci-fi fantasies though.

mlinsey··on Governance of Superintelligence
In the scenario where the current AI boom takes us all the way to AGI in the next decade, IMO there is little downside. Risks are very large, OpenAI/Sam have expertise, and their novel corporate structure, while far from completely-removing themselves from self-centered motives, sounds better than a typical VC funded startup that has to turn a huge profit in X years.

In the scenario where the current wave fizzles out and we have another AI winter, one risk is that we'll be left with a big regulatory apparatus that makes the next wave of innovations, the one that might actually get us all the way to an algined-AGI utopia, near-impossible. And the regulatory apparatus will now be shaped by an org with ties to the current AI wave (imagine the Department of AI Safety was currently staffed by people trained/invested in Expert Systems or some old-school paradigm).

mlinsey··on Google DeepMind
Drawing a bright line between research vs products and considering only the former to be "innovation" is a way to become quickly irrelevant.
mlinsey··on Yann LeCun and Andrew Ng: Why the 6-Month AI Pause Is a Bad Idea [video]
> I’m seriously confused why anybody thinks anyone will even remotely care about any kind of pause is asked for.

I can say that Google and Microsoft would definitely care, just as two specific instances of "anybody" that I have worked for and seen what their legal review process gating new product launches is like.

You seem much more optimistic that the community efforts are <6 months behind those companies than I am; I would be happy to be wrong here and see an open-source self-hosted LLM with better than GPT4 performance by September.

mlinsey··on Yann LeCun and Andrew Ng: Why the 6-Month AI Pause Is a Bad Idea [video]
GPT4 is expensive enough to train that most organizations with that high of a compute budget also have compliance departments and will not flat-out ignore a law like that. State actors overseas and the indie hackers with open-source models will not be deterred, but they are probably six months behind anyway. I don't think the six-month moratorium (as opposed to say, a three year moratorium) would be so impractical to implement, for that reason. But I'm also not sure what it would get us, the past six months of alignment work doesn't suggest six more months would solve the problem.
mlinsey··on We need a more sophisticated debate about AI
I'm not convinced we understand qualia/feelings well enough to definitively say. The presence of eg phantom limb pain suggests pain can arise in the mind, independent of physical nerve signals. It would be a remarkable convergence if AGI's developed the same sensations as humans, but we can't rule it out until we fully understand both how the human mind and the AGI systems actually work.
mlinsey··on Reddit has been down for more than an hour
I think it's downvoting. It's one thing to see (what you think is) a really bad reply to you get some likes, it's another to see a negative number next to your comment, I feel much more compulsion to defend myself for the latter.
mlinsey··on Why Not Mars
the "eggs in one basket" argument only makes sense if Martian colonization is itself just step 1 on a list of dozens of similarly-difficult leaps to get us to terraforming and interstellar colonization.

It's a very exciting, multi-generational project to think about. Though frankly I think we're going to get AGI + brain/computer interfaces + cloud mind uploads before we get martian terraforming; and dropping the bag-of-meat related requirements would make extraterrestrial colonization much easier.

mlinsey··on How Southwest Airlines melted down
Mobile apps or a web portal absolutely could have helped, simply by massively increasing the throughput of crew members being able to report that they didn't make it to their assigned location. Which could have stopped the cascading effect much sooner.

The manual "reporting/fixing" of issues, which was done over the phone, includes pilots and crew members reporting "hey, I didn't actually get on the flight I was scheduled to". If the system did not receive that message, it was built to assume that the crew members made it to their next location and would be available for their next scheduled flight. But once a certain threshold of flights got canceled because of the storm, the limiting factor became how quickly staff members could even report that they didn't make it to their next location - they'd stay on hold for hours, until their next flight in their next city was scheduled to leave, and the system would still assume they were on that flight because it hadn't yet heard otherwise.

Lack of webapps certainly wasn't the only issue here - as the article notes, their whole point-to-point scheduling model is more vulnerable to cascading failure as well, vs. a hub-and-spoke model. But the phone-based system did certainly cause a feedback loop that kept the disruption happening for much longer than it should have. And moving away from point-to-point would have many more tradeoffs/downsides for consumers than simply modernizing their tech system.

mlinsey··on The hype around esports is fading as investors and sponsors dry up
As far as I can tell, two groups are making plenty of money off of esports:

1 - The companies that make the game. Whether it's Riot and Blizzard selling slots in their leagues for eight figures, Valve using their annual tournament to sell in-game cosmetics, or all the companies ultimately owning broadcast rights to their game, this is the biggest difference between esports and traditional sports.

The New York Yankees and Los Angeles Dodgers make money because they sell their own tickets and broadcast rights to their games; even though MLB does control the streaming revenue, they share it out to teams and ultimately have to bow to the wishes of a majority of team owners. What you don't see is a single company MLB inc that owns the copyright to the game of baseball, sells broadcast rights to all teams themselves, charges teams to play in the league, and can kick teams out of the league at a whim. That's the situation in esports.

2 - Individual players, with streaming. Players can first make a name for themselves in competitions, then stream on their own Twitch channel for revenue. This is not that different from athletes acting as social media influencers and signing endorsement deals, but the biggest difference is that by streaming on Twitch, they appear side-by-side with tournament broadcasts. It's as if LeBron's instagram account where he streamed his workouts and pickup games were just one change-of-the-channel away from ESPN, and people would consider it normal to flip between the game and individual player streams.

Lots of esports orgs, as part of signing players, get a big cut of the player's streaming revenue. But the revenue for an individual player's twitch stream, while great for an individual, usually isn't going to be significant enough to maintain a whole organization, and when a player brand does get big enough that their stream could sustain an org - that's when the player will be heavily incentivized to go independent, and make more money from streams than they do from competing.

Ultimately, I think esports has a bright future - overall total viewership continues to rise, even though some games like League of Legends - which is more than a decade old now - are starting to fade. It's just the business models of the offline sports world don't carry over, and that's especially apparent with these organized teams.

mlinsey··on U.S. News pulls Columbia University from its 2022 rankings
This is why it is really just Harvard (likely Stanford, Yale, and Princeton too, but definitely Harvard) that can make a statement boycotting the list, just like a billionaire can get away with wearing a plastic $10 watch but mere millionaires feel the need to wear a Rolex to show off their money. In fact, it wouldn't surprise me if Harvard did the least to game the US News metrics of any top 25 university, because the US News staff would just tweak their formula to keep Harvard in the top 5 regardless.
mlinsey··on Twitter set to accept Musk's $43B offer – sources
It's 2,000 shareholders, and you are allowed to go above the limit, it's just that if you do, then you have to make public financial statements and in general follow all the disclosure rules and reporting requirements of a public company. Since it can be expensive to follow these requirements, especially for a startup that's not setup to do those sort of disclosures, startups will carefully avoid going over the limit. But since Twitter has already been public up until now, all the processes and institutional know-how to comply with those requirements is already in place, so continuing to comply shouldn't be too hard.
mlinsey··on The YC Winter 2022 Batch
I think in the scenario where miners block the move to PoS (or any other much-more-scalable technology) is a scenario where nobody ever actually uses crypto for anything, and almost all coins go to $0 after the speculative bubble pops. To be clear, I think that's certainly a possible world, maybe even the modal outcome.

The alternative world, where we're using crypto to pay for everything, and we're no longer logging into twitter, but instead using our keys to log into a client for a globally distributed twitter where we own all our data on a blockchain - that's impossible without not just Eth2 but a several orders of magnitude of improvements beyond that.

Basically, I think taking the current per-transaction carbon impacts, multiplying it by eg. Visa's transaction volume, and forecasting a world where crypto is burning 1,000X as much carbon as the whole transportation sector or whatever - is a fallacy. Crypto will either solve the scalability problem (and therefore consume a much more moderate amount of energy), or it won't be used at all.

← PreviousPage 3 of 22Next →