ericjang

It strikes me as deeply inauthentic that people can be so passionate about some social issues and utterly indifferent or complicit to others. I understand that people can prioritize different issues at different times, and some people say that it's even possible to care about one issue without the expense of another.

Nonetheless, that statement is not substantiated by the level of concern U.S. citizens and woke-minded folks around the world devote to issues like the soft economic power China wields over SouthEast asian countries, the oppression of Uyghur minorities in China, or the political status of pro-democracy countries like Taiwan (to name a few).

Do people only care about a political issue if people threaten to cancel them on Twitter?

[1] https://twitter.com/zoom_us/status/1266877451646283776?s=20

This is a topic with a lot of nuance and I'm no Rowling when it comes to writing or discussion about gender and sex. I think she's making clear that she's a feminist first, and that she's taking issue with how any woman who does not fall in line with the "approved hashtags" regarding trans issues is labeled a TERF. All nuance on what she actually believes (3690-word essay) has been replaced by a catchy, provocative statement like "you are killing trans people with your ideas."

It would be so much easier to tweet the approved hashtags – because of course trans rights are human rights and of course trans lives matter – scoop up the woke cookies and bask in a virtue-signalling afterglow

However, an indexing strategy still requires an estimate of the value of coronavirus research. Suppose $5T USD was funneled into all research over the last 10 years. How much of that would virology/epidemiology have gotten, and would it have been enough?

I think this only works with the caveat that the core code is so well-designed that the user is capable of mixing contrib APIs with core APIs without requiring action on part of core developers.

For instance, TensorFlow v1 had a contrib directory that encouraged a lot of internal contributions and features to be added. This was initially quite nice due to the "batteries included" argument. However, the success of TF quickly resulted in a lot of bloat being added. There were competing ways to do the same thing within the default TF installation, and the worst part was that users expected contrib features to be interoperable with all other contrib features and core TF.

1) "It’s another big jump in the number, but the underlying architecture hasn’t changed much... it’s pretty annoying and misleading to call it “GPT-3.” GPT-2 was (arguably) a fundamental advance, because it demonstrated the power of way bigger transformers when people didn’t know about that power. Now everyone knows, so it’s the furthest thing from a fundamental advance."

2) "The “zero-shot” learning they demonstrated in the paper – stuff like “adding tl;dr after a text and treating GPT-2′s continuation thereafter as a ‘summary’” – were weird and goofy and not the way anyone would want to do these things in practice... They do better with one task example than zero (the GPT-2 paper used zero), but otherwise it’s a pretty flat line; evidently there is not too much progressive “learning as you go” here."

3) "Coercing it to do well on standard benchmarks was valuable (to me) only as a flamboyant, semi-comedic way of pointing this out, kind of like showing off one’s artistic talent by painting (but not painting especially well) with just one’s non-dominant hand."

4) "On Abstract reasoning..So, if we’re mostly seeing #1 here, this is not a good demo of few-shot learning the way the authors think it is."

--------- My response:

1) The fact that we can get so much improvement out of something so "mundane" should be cause for celebration, rather than disappointment. It means that we have found general methods that scale well and a straightforward recipe for brute-forcing our way through solutions we haven't solved before.

At this point it becomes not a question of possibility, but of engineering investment. Isn't that the dream of an AI researcher? To find something that works so well you can stop ``innovating'' on the math stuff?

2) Are we reading the same plot? I see an improvement after >16 shot.

I believe the point of that setup is to illustrate the fact that any model trained to make sequential decisions can be regarded as "learning to learn", because the arbitrary computation in between sequential decisions can incorporate "adaptive feedback". It blurs the semantics between "task learning" and "instance learning"

3) This is a fair point actually, and perhaps now that models are doing better (no thanks to people who spurn big compute), we should propose better metrics to capture general language understanding.

4) It's certainly possible, but you come off as pretty confident for someone who hasn't tried running the model and trying to test its abilities.

Who is the author, anyway? Are they capable of building systems like GPT-3?

A separate set of 3 parameters also parameterizes the SIR compartmental model (https://en.wikipedia.org/wiki/Compartmental_models_in_epidem...), the solution of which also looks like a logistic curve. But this is a model-based (dynamics-based) solution whereas one may be interested in just fitting the logistic model based on the assumption that it's going to be logistic.

[1] https://en.wikipedia.org/wiki/Logistic_function

Assumption: you use a separate computer for your personal computing needs not related to work.

Disclaimer: I'm not a doctor or an expert.

Chloroquine is suspected to have general prophylactic abilities (like Remdesivir), but also hypothesized that if taken with Zinc, [1] acts as an ionophore with which Zn2+ ions can inactivate the first part of the viral replication step. But at this time, clinical trials for HCQ have been somewhat mixed, with small N-size studies reporting weak benefits or lack of benefits. I don't know of any RCTs that include Zn2+ in the treatment group along with HCQ. It's still as yet unclear.

[1] https://www.youtube.com/watch?v=U7F1cnWup9M

Pretend for a second that you're Andrew Cuomo on January 1, 2020. What information would you have to justify this crash program?

Pretend that you're Andrew Cuomo on April 4, 2020. You're aware of what stage China is in. What would you do differently?

Would you have instituted a statewide lockdown when there were 1-2 cases in California?

Would you be willing to trade certain unemployment and economic unproductivity for 2+ months, in exchange for the possibility that you might end up like Italy instead of Taiwan/South Korea? The last datapoint we had was SARS-Cov-1, which ended up not becoming a huge pandemic. Of course, the disease characteristics for SARS-Cov-2 are different, but without the benefit of hindsight, was the decision really that clear?

Incompetent US administration aside, the financial markets - which can be thought of as an expensive forecasting computer - certainly did not price in what would happen, so I'm not sure we can expect policymakers to arrive at far-superior decision making.

Considering that China is not completely out of the woods yet either, and the US/Italy lags about 1 month behind them in pandemic progression, how would you advise the US act now, with the knowledge of what China is doing?

I agree with your observation of what happens to the black and white billiards. My question is - if the interactions are all time-symmetric (implies conservation of entropy), where does the entropy come from in the macroscopic system? My belief is that the microscopic model is flawed; if we simply introduce some randomness into the collision dynamics of two particles (which is believable given particle accelerator experiments / QM), then microscopic interactions are no longer T-symmetric.

It comes down to a philosophical interpretation of whether you can treat a system of two particles probabilistically (i.e. thermodynamically in aggregate) or not.

If we assume molecular chaos hypothesis (independent particle velocities prior to collision) and take into account probabilistic collisions occurring at the 2-particle microscopic level, then post-collision velocities are now dependent, thereby imposing an ordering to time.

I have to think more about this though, and I'm still struggling with understanding the circular dependency between entropy "being time" and entropy being "caused" by "motion" (with respect to ... time?)

What is a "stable grasp"? An entire body of classical robotics literature is devoted to answering this problem. I work on robotic grasping research and defining what it means to grasp something successfully is a little tricky unless you have a broader context of what the grasping is for.

Nature doesn't care whether hands exist or not - only that the animal can feed. That is why we see such a spectrum of appendages in the animal kingdom. Some are hand-like, others not at all, and some in between. In some sense, "hand" is a mental construct of human language.

If thought is indeed "intrinsically dependent on biochemistry...", then we have a hard journey ahead.

However, AI philosophers should try their own medicine and present claims with real contingencies, i.e. make claims that are experimentally falsifiable.

Put another way, I'm less interested in debating what "understanding" is and more interested in figuring out how to build a Chinese room to begin with. Once we have such a system, perhaps we might realize that the Chinese Room thought experiment was ill-posed to begin with.

He's making the point that the ML field's obsession with causal inference (and causal discovery) is overrated, precisely because our gold standards of interpretable, safety-critical systems (airplane flight) are based on Navier Stokes/CFD. Planes were made to fly based on empirical validation of these models, long before we gained a more detailed understanding of how causality (the equations and models themselves are time reversible, implying that they contain imperfect knowledge of causality)

The same goes for Apple and Airpods.

I think the title and lede, "working for Microsoft cost me $200M" is intended to be provocative and encourage clicks on the article, while providing an interesting example of a "road not taken" in Garry's professional life. The dollar magnitude is unusual, but the feeling of regret is relatable to us all, even if you took the opposite bet that Garry did (e.g. turning down a LinkedIn or Coinbase offer in 2016 to work at a failed startup instead).

Since this article is on the top of HN today, I'd say Garry succeeded at sharing his personal anecdote.

On the content itself (startup vs. bigco), he is right about "you only can make a fraction of the value you create if you don't own equity" and "If you don't work on your dreams, someone will put you to work on theirs." But then it follows that everyone should start their own company if they are capable, or work for a bigco if they are not. The current SV labor market reflects this reality as well.

It's possible to sell some top-tier talent on taking less than 1% equity to work on your dreams, but they must really believe in the founder's vision and the founder's ability to provide a working environment that complements what the employee is not able to de-risk by themselves.