Isinlor

There is a prediction market called Metaculus.

TL;DR In 2020 community of 169 people and the best forecasters were assigning ~15% that it will happen by July 2021.

More specifically, on Dec 31, 2016 in partnership with Center for the Study of Existential Risk, Machine Intelligence Research Institute, and The Future of Life Institute they asked:

How long until a machine-learning system can take a simple text description and turn it into a program coded in C/Python?

https://www.metaculus.com/questions/405/when-will-programs-w...

First 19 forecasters in March 2017 were predicting mid-2021, the best forecasters were predicting late 2024. When the question closed in 2020 the community was predicting January 2027 and the best forecasters were predicting March 2030.

The question resolved on July 2021 when Codex was published.

Community and the best forecasters were assigning ~15% that it will happen by July 2021.

I'm currently 14th best forecaster there and I was predicting 33% before July 2021. It was my last prediction, and it was made on October 2018.

I'm also predicting 75% that we will have AGI by 2040 as defined in this question:

https://www.metaculus.com/questions/3479/when-will-the-first...

20% that it will happen before 2030.

There is also stronger operationalization:

https://www.metaculus.com/questions/5121/when-will-the-first...

My prediction here is 60% before 2040 and 5% before 2030.

I have also "canary in the coal mine" questions:

When will AI achieve competency on multi-choice questions across diverse fields of expertise? Community predicts 50% before 2030, I agree.

https://www.metaculus.com/questions/5276/ai-competence-in-di...

When will AI be able to learn to play Montezuma's Revenge in less than 30 min? Community predicts 50% before 2025, I think 50% before 2027.

https://www.metaculus.com/questions/5460/ai-rapidly-learning...

What does Yahoo Japan have to do with Europe?

There should be some separate law that would oblige companies to accept "do-not-track" HTTP header.

Then we could just set it in our browser settings.

It's not about Google or Facebook.

It is about USA politicians and USA laws that allow USA to do anything with our data as if we were terrorists.

The "Privacy Shield" was established by politicians twice already and dismantled twice by EU courts.

The initiative to strike down privacy shield and enforce ban is a result of civil society.

https://noyb.eu/en/austrian-dsb-eu-us-data-transfers-google-...

> 2020 CJEU ruling hits the real world. In July 2020, the CJEU has issued its groundbreaking "Schrems II" ruling, holding that a transfer to US providers that fall under FISA 702 and EO 12.333 violate the rules on international data transfers in the GDPR. The CJEU consequently annulled the transfer deal "Privacy Shield", after annulling the previous deal "Safe Harbor" in 2015. While this sent shock waves through the tech industry, US providers and EU data exporters have largely ignored the case. Just like Microsoft, Facebook or Amazon, Google has relied on so-called "Standard Contract Clauses" to continue data transfers and calm its European business partners.

> Max Schrems, honorary chair of noyb.eu: "Instead of actually adapting services to be GDPR compliant, US companies have tried to simply add some text to their privacy policies and ignore the Court of Justice. Many EU companies have followed the lead instead of switching to legal options."

Yes, Texas does not allow direct sale of cars.

You have to buy your Tesla out of the state and Tesla will deliver the car to Texas.

https://cbsaustin.com/news/local/texas-law-keeps-teslas-made...

https://insideevs.com/news/510302/tesla-no-sales-allowed-tex...

Biden administration is explicitly designing laws to prop up other car companies and exclude Tesla.

It's not just about labor PR, they want to force Tesla to unionize.

But Tesla will not unionize and they will be the only ones without subsidies. Extremely shitty lobbing.

https://www.cnbc.com/2021/10/29/energy-secretary-defends-tes...

He does not have good lobbyists.

Have you seen Tesla being the only company not invited on electric car day in white house?

https://www.autoweek.com/news/green-cars/a37244461/elon-musk...

Do you know that Tesla will be excluded from tax credits for EVs?

https://www.cnbc.com/2021/10/29/energy-secretary-defends-tes...

Do you know that Texas where Tesla is building its next factory does not allow to sell Teslas?

https://cbsaustin.com/news/local/texas-law-keeps-teslas-made...

This clearly falls into Art. 6 GDPR paragraph 1, point b) :

Processing shall be lawful only if and to the extent that at least one of the following applies:

b) processing is necessary for the performance of a contract to which the data subject is party or in order to take steps at the request of the data subject prior to entering into a contract;

This legal basis is a lot more clear and a lot less stringent than point f) legitimate interest as it does not explicitly require you to establish "legitimacy" and balance it against vague "interests or fundamental rights and freedoms of the data subject".

f) processing is necessary for the purposes of the legitimate interests pursued by the controller or by a third party, except where such interests are overridden by the interests or fundamental rights and freedoms of the data subject which require protection of personal data, in particular where the data subject is a child.

See: https://gdpr-info.eu/art-6-gdpr/

When you do anything with personal data in the EU you have to have a legal basis.

There are 6 and only 6 possible legal basis: https://gdpr-info.eu/art-6-gdpr/

But most businesses will be choosing from:

- the data subject has given consent to the processing of his or her personal data for one or more specific purposes;

- processing is necessary for the performance of a contract to which the data subject is party or in order to take steps at the request of the data subject prior to entering into a contract;

- processing is necessary for compliance with a legal obligation to which the controller is subject;

- processing is necessary for the purposes of the legitimate interests pursued by the controller or by a third party, except where such interests are overridden by the interests or fundamental rights and freedoms of the data subject which require protection of personal data, in particular where the data subject is a child.

You can do basically anything with things like IP addresses as long as you have valid consent from the client i.e. they need to actually know, or at at least be able to learn, what you are doing with their data and decide that it is ok. So, no guessing here, just be transparent, and assume no consent by default.

In case of ISP they have to process your personal data because it is necessary for the performance of a contract of providing the internet service. Also, no guessing here.

The legitimate interest clause is a "catch all" clause for anything that legislator did not think about, so it is very vague by design. You do not want to choose this as a legal basis for data processing if you do not want to deal with legal uncertainty. But if you do choose it, you should have strong arguments that you really need this legal basis.

If similar companies to yours are able to do exactly the same thing in a way that is less impactful on privacy then you can expect that courts will not grant you a legitimate interest.

You can also do legal tests do determine whether you have a legitimate interest:

- The purpose test (identify the legitimate interest);

- The necessity test (consider if the processing is necessary); and

- The balancing test (consider the individual’s interests).

See more detail here: https://ico.org.uk/for-organisations/guide-to-data-protectio...

Also, based on my observation if you are not doing anything really egregious and you are willing to cooperate with data protection agencies (DPA) you do not have to worry about anything. If DPA decides you are doing something wrong they will tell you about it. And if you just adjust, like start to host fonts on your servers, they will let it slide or give you a small slap on the wrists. The really high fines are reserved for malicious conduct or gross incompetence with actual harm already done to people.

This ruling will 100% be upheld in the higher courts.

The website is arguing that they have a legitimate interest in downloading fonts from Google in client browser, but as the court correctly states the website can provide these fonts directly. There is no reason to infringe on the user privacy, so there is no legitimate interest. And therefore use of Google fonts was without a legal basis.

BTW - The website could have used a different legal basis out of 6 available, like consent. See: https://gdpr-info.eu/art-6-gdpr/

> I'm curious to know whether DNS and your IP being in the the header of packets travelling through various different countries that can be sniffed is also considered as unwilful data sharing?

Unless there is another way to achieve the same purpose there is a legitimate interest in processing that data for the purposes expected by the client i.e. providing internet service.

Governments should commit to creating this type of SuperClusters designed for AI workloads with the latest technology every 2-3 years for academic use.

It is really peanuts for US government or EU.

I went trough Quantum Mechanics course and it really helped me understand Quantum Mechanics on mathematical level. I had my head full of Quantum woo from popular science programs and I barely could make heads or tails out of it. Knowing basic mathematics behind it helped a lot.

PHP is share-nothing architecture. Nobody really uses threads in mainline PHP.

Whole class of problems around multi-threading, mutexes, dead locks etc. are really foreign to PHP developers.

I'm not sure if you did not read what I quoted, or I misunderstand something.

Yes, PHP has reference counting, but it also has specific process that kicks in to clean up circular references.

Share-nothing architecture, good, stable ecosystem build around Composer, Laravel and Symfony and as you said a lot of devs.

Otherwise, I think differences in syntax etc. are rather subjective and everyone has their preferences. For example, I really like PHP "arrays" because it removes cognitive load of choosing between different data structures. When writing business rules 90% of time I don't really care about optimization or memory, but about readability and communication.

Overall, PHP is rather dumb language and I love it for being dumb.

Python package management currently is a nightmare, but projects like Poetry seem on a good track to make it better. JavaScript is better than Python at package management, but it has issues with many small packages maintained by random people. Java ecosystem also seems quite a lot more complex, especially when you compare Composer and Maven. Compilation aspect also changes the way you work.

I have no experience with Ruby, C#, .NET etc.

There is a garbage collection:

https://www.php.net/manual/en/features.gc.php

> This section explains the merits of the new Garbage Collection (also known as GC) mechanism that is part of PHP 5.3.

> First of all, the whole reason for implementing the garbage collection mechanism is to reduce memory usage by cleaning up circular-referenced variables as soon as the prerequisites are fulfilled. In PHP's implementation, this happens as soon as the root-buffer is full, or when the function gc_collect_cycles() is called. In the graph below, we display the memory usage of the script below, in both PHP 5.2 and PHP 5.3, excluding the base memory that PHP itself uses when starting up.

I mentioned ARC because based on your description it should be "the kind of problem that ILP eats for breakfast". But based on what you say about required effort it does not appear to eat it for breakfast and requires significant engineering fine-tuning, so I call foul on your previous description of your work.

I agree, to a degree, with your assessment of the problem of cheating. Currently community is applying two fixes to that: 1) require few-shot learning ability 2) require good performance on many different few-shot learning benchmarks

My hobby is future forecasting and I was thinking for some time about creating a benchmark for AGI based on Bongard problems, ARC and Montezuma's Revenge. I did have a question about ARC that managed to surprise community: https://www.metaculus.com/questions/3762/will-early-2020-ai-... I have one about Montezuma's Revenge where I require some reasonable score after 30 min of game play, no specific to Montezuma's Revenge engineering allowed, not even retrieving rewards from internal game memory: https://www.metaculus.com/questions/5460/ai-rapidly-learning... I'm aware of the paper you mentioned around ARC and my fix to the benchmark would be to require few-shot generalization on task-level e.g. tested system, if it is a learning system, must have access only to couple of tasks for training before start of evaluation. The same with regard to Bongard problems. Requiring that all 3 benchmarks are solvable by one unified system IMO would be a reasonable test for general, visual intelligence.

DeepMind recently decided to include comparison of Gopher performance on Massive Multitask Language Understanding benchmarks vs. forecasters prediction. Forecasting communities seems to have gained some clout in AI circles (OpenAI is also cooperating with Metaculus on AI forecasting). Few-shot ARC forecast from Metaculus could give you an external estimation of the difficulty that you could use against reviewer 2.

Also, despite what you think about large language models memorizing everything few-shot SuperGlue is still standing: https://www.metaculus.com/questions/4932/when-will-ai-achiev... My intuitive evaluation of GPT-3 light text reasoning skills based on playing around with it matches few-shot SuperGlue score of 72 vs. 90 for avg. human.

> Meanwhile, some big, 30-person team at a large tech corp with a few milllion dollars' budget will be throwing a gigaflop of compute and a few terrabytes of data on the problem and "solving" it without having to learn any "core priors", by finding some way around the dearth of training data, just as was done with Bongard problems and Winograd schemas.

Compute is cheap and it reliably gets cheaper. Models like GPT-3 even with cost of tens of millions of dollars are economically justifiable. It's quite possible that OpenAI is currently turning profit on GPT-3. Real market performance is the ultimate benchmark of whether "performance degrades precipitously once exposed to real-world data".

I'm not a biochemist. I don't know how much knowledge of how protein fold is important as opposed to knowing the final structure. My understanding is that current expensive method of X-ray crystallography in the range of $100k per protein also does not explain how protein fold, however the structure itself appear to be useful.

Many complex problems do not appear to be explainable in the way planet motions are explainable. The fact that many aspects of the universe are explainable in simple laws is rather extraordinary in itself. In reality, there may be just a lot of irreducible complexity. Maybe success of Deep Learning teaches us that certain things do not fit into framework of science that you assume to be the only correct framework. Reality does not care about what we would want to be true.

Maybe invention of AlphaFold will lead to better understanding of protein folding the same way that invention of steam engine lead to the laws of thermodynamics. Historically, science often lags behind progress of technology. Time will judge.

I also thought that the lack of resources will be your answer. In 2004-2009 neural networks were in quite similar position to ILP. However, NN showed significant potential as evaluated by the broad community at the time and so resources allocated to NN started to grow exponentially. ILP and symbolic methods more broadly do not show similarly convincing potential to the broad community. I'm sure OpenAI, DeepMind, Google, Micosoft etc. would not hesitate to dump equivalent resources they do into many other projects if they saw potential. DeepMind, Microsoft and Google especially hire people with very broad range of expertise.

What may be true is that the world is starting to reach limits of the resources that can be dedicated to deep learning. Exponential growth in resource use can not go forever unless deep learning will start feeding back into available resources and we are certainly not there, not yet.

> Yet the algorithms that I study run on a student laptop, are trained in a weakly supervised manner, take seconds to train, and generalise robustly from single examples without any sort of pre-training whatsoever

Show me that applied to: https://github.com/fchollet/ARC

1) I have access to OpenAI Davinci-Codex. I can and I did recreate their results on few prompts.

2) It is not really an issue of how easy it is to break. If it is able to solve the problem it will do it somewhat reliably. The issue is that questions provide solutions. So, "Solve each equation for x: ln(x*2-1)=3" will just keep on generating more similar problems, but "Using Sympy, solve Eq ln(x*2-1)=3 for x." will work.

For example, you will never get this prompt to work: "Find the limits as x → ∞ and as x → −∞ . Use this information, together with intercepts, to give a rough sketch of the graph as in Example 12. y = x 2 (x 2 − 1) 2 (x + 2)" but their step by step what to do prompt does work.

If you don't know Python ecosystem and you don't know how the solution should look like then you are out of luck.

3) Sympy does not seem hard to use. Codex is actually pretty good at writing "template" code. It struggles with anything that requires some level of reasoning instead of matching template to a problem.

4) Could you point me to the exact questions? I can try to recreate it.

OpenAI Codex is actually quite good. It is groundbreaking in comparison to anything before it, but it is far, far, far from perfect. It's like really bad stackoverflow in the sense that if you probe it couple of times it will provide you reasonably looking solution that may or may not work.

I would also not focus on the current state. Trajectory of improvement is more important. 5 years ago you could not get anything even comparable to Codex, in next 5 years it may just be far from perfect, and in 10 years it may actually work. On many metrics, on average over years, Deep Learning is improving 2x every year or so.

> and then fine-tuned again on the kinds of problems that it is supposed to solve

Could you point me to where they claim to have fine-tuned Codex?

From what I can see they claim:

> We use OpenAI’s davinci-codex engine for all of our generations. We fix all of Codex’s hyperparameters to be the same for all experiments: top-p which is the portion p of the token probability mass a language model samples from at each step is set to 1, sampling temperature is set to 0 (i.e. argmax), and response length is set to 200 tokens.

BTW - OpenAI Davinci costs $0.06 per 1000 tokens. Codex is currently free in closed beta, but I guess cost will be the same. I would be happy to pay even 100x ($6) for correct solutions to advanced mathematical problems. The issue is that this paper has ridiculous evaluation and Codex does not work anywhere near as good as they claim. It is pure hype by people who do not appear to be affiliated with OpenAI.

They are using OpenAI Davinci-Codex, comparable to GPT3-175B, without any further fine-tuning as far as I can tell. Over-fitting, in the traditional sense, specifically to these questions is very unlikely.

What is problematic is at the bottom of page 6:

> Prompts that result in correct solutions, whether from original or modified prompts, are used for evaluation metrics.

In other words, they reject all results without correct solutions and claim perfect accuracy...

And as you noticed even their evaluation is really not good enough to establish perfect accuracy.

>> That is an example of the difference between expensive toys like Dall-E and world-changing scientific work.

There isn't really that much difference. DeepMind took some 3 years to max out CASP protein-folding benchmark running from 1994 on a problem known since ~1960. Protein folding is exactly the type of world-changing scientific work. But it is just a tip of the iceberg. DL is being successfully applied to PDE solving or electron density predictions and more.

The psychedelic hamster dragon is an example of the fact that DL is capable of 0-shot generalization. It is a significant scientific observation in itself.

And with regard to your example. We are discussing whether deep learning has staled. I give you mostly examples from last year like EfficientZero (30 Oct) ~500x improvement in sample efficiency over 2013 DQN, some literally not older than a month: GLIDE (20 Dec), Player of Games (6 Dec).

You are giving me example from 2004 (2009?).

I'm not trying to troll you. I just think you are biased in you assessment of significant results and what it means to be stalled.

So much that you write is blatantly false, or stretch to the point of being as good as false.

But you seem well read, so I don't think there is a point in trying to convince you.

It's also interesting coming from someone researching Inductive Logic Programming that provides no results at all.

From my perspective, I don't care for theoretical advances. I care about results and if scale provides results, I'm happy.

For people who are just reading this exchange take a look here:

https://pbs.twimg.com/media/FHHPjU0VIAUrSMA?format=jpg&name=...

It's 16 computer generated images from the description below each image. Having capability like that is just very useful and fun.

I wish you also luck in finding a "cat paying checkers in style of Salvador Dali" or a "psychedelic painting of a hamster dragon" on the internet.

There is a lot of stuff on the web, but not everything.

How has the progress stalled?

Just last year there was a breakthrough in decades old problem of protein-folding. The year before that we figured out that large language models have astonishing few shot learning abilities. CLIP introduced strong 0-shot capabilities to image recognition. Progress in image generation from descriptions has been nothing but astonishing with Dall-E and GLIDE from OpenAI.

Not to mention breakthroughs in game theoretical problems like Counterfactual Regret Minimization and more recent application of Deep Learning to make CFR practical in actual games. Including continuous progress in the AlphaGo linage to AlphaZero, MuZero, and last year EfficientZero and Player of Games.

It's weird how many unrecognized geniuses there are in the neuroscience all claiming to have explained the brain and people just need to recognize their genius. All sounds very similar to Jeff Hawkins.

There is very little theory to DL. It's almost completely about empirical results.

You will be very disappointed if you expect that there is some grand theory of Deep Learning.

My suggestion is to follow Arxiv Sanity: http://www.arxiv-sanity.com/top

Sam Altman, Jul 19, 2020:

> The GPT-3 hype is way too much. It’s impressive (thanks for the nice compliments!) but it still has serious weaknesses and sometimes makes very silly mistakes. AI is going to change the world, but GPT-3 is just a very early glimpse. We have a lot still to figure out.

Slavic people do not belong to Russia in any sense either. Certainly I do not.

Even Russian speakers do not always want to associated themselves with Russia.

Here is a primary school in winter, in Oulu, Finland, close to Arctic circle:

https://www.youtube.com/watch?v=Uhx-26GfCBU&t=72s

200 000 people live there, and in January, during polar twilight their daily average temperature is −8C (17F) and 50% school trips are on bikes.

I guess I formulated it in a confusing way. I don't know whether myocarditis caused by vaccine is more or less severe vs. myocarditis caused by COVID-19.

What the article references is a report that 76% of vaccine-associated myocarditis events were described as mild [0].

> A total of 76% of cases of myocarditis were described as mild and 22% as intermediate; 1 case was associated with cardiogenic shock. After a median follow-up of 83 days after the onset of myocarditis, 1 patient had been readmitted to the hospital, and 1 had died of an unknown cause after discharge. Of 14 patients who had left ventricular dysfunction on echocardiography during admission, 10 still had such dysfunction at the time of hospital discharge. Of these patients, 5 underwent subsequent testing that revealed normal heart function.

We also know that long term effects of COVID-19 overall i.e. not only myocarditis are not always mild and can be debilitating even for people below 40, even after mild COVID-19. See my reply here: https://news.ycombinator.com/item?id=29564815

[0] https://www.nejm.org/doi/10.1056/NEJMoa2110737

Indeed, they can not know quantitatively lifetime risk of COVID-19 after 24 months of observation. But we do know from clinical experience that COVID-19 is damaging a lot of people in very severe ways. I was also quoting data on long COVID that indicate that a lot of people (10% and more) suffer long COVID after 3 and even 6 months.

We also have some data on SARS survivors [0]. The SARS virus is closely related to SARS-CoV-2, but it is a lot more severe.

> Dr. Paula Gardner, a psychologist at St. John’s Rehab in Toronto, has been documenting the health of survivors, post-SARS.

> “They’re still, after 10 years, experiencing problems. Issues such as fatigue, muscle and joint pain, shortness of breath and some newly developing problems such as neuropathy, numbness in the feet and hands,” she said.

We have also experience with the general chronic fatigue syndrome that seems to be related to infectious diseases: https://en.wikipedia.org/wiki/Chronic_fatigue_syndrome#Viral...

[0] https://globalnews.ca/news/404562/sars-10-years-later-how-ar...