Do you mean the "internal monolog" in parentheses are added by you, and only the text after the parentheses are generated by GPT-3? Either way, it's kind of fun to think of it as kind of an experiment on the bicameral mind theory (or at least Westworld's interpretation it), with the "monolog" being provided by you, and GPT-3 responding to the "internal voice from "God". Maybe that's how GPT-4 can gain consciousness, or at least the appearance of gaining it :)
Cargo-Cult programming is already happening at a not-insignificant scale due to the vast amount of sample code/tutorial/Q&A online. There could be lots of nuances in that snippet you just copied from a SO comment, which "just works". GPT-3 based tools could help you generate a working (or barely working) sample fairly easily. But the decision to just ship that code as is, or to understand it and tune it for your specific needs still mostly depends on the developer.
This, the difference between one-shot fine-tuning, vs fine tuning for GPT-2, is one of the major breakthroughs. Since GPT-3 is so hot in the past few days, people seem to forgot or not realize lots of the GPT-3 examples shown off today were possible with GPT-2, with the catch that you had to fine-tune your own GPT-2 model to fit your problem domain (game plots, poems, music, bots that chats like certain characters, etc). GPT-3 makes that fine tuning process unnecessary (although practically you probably can't/can't afford to fine-tune your GPT-3 model)
We fine-tuned GPT-2's 774M model using all seasons' South Park scripts. The result is a model that we can use to generate chats/conversations between South Park characters. If you've watched the show before, the characters' language style is extremely on point w.r.t. who they are in the show. Depending on the initial message fed into the conversation (either random, or from a live tweet) the conversation could sometimes be hilarious. Let us know what you think!
Exactly! I didn't know the author, but couldn't help but check on his background after reading the blogpost.
It's almost a dead giveaway for excerpts like 'The notion of a counter is not that of a single concrete class of object, it is an abstract description of the behaviour of a system providing certain operations satisfying certain rules.' , and 'A number is not any one thing, it is any one of any number of things that implement some operations (and there are different types of number depending on what operations you want to implement).'
FWIW, "Amazon EC2 Queue" actually covers a lot of real world use cases for SWF where a "queue" is needed to manage states and distribute works among workers.
The list is a bit out dated, and as for Amazon State Machines, it's a name more suited for AWS Step Functions, which is the successor of SWF (and defined as an actual state machine). Step Functions do cover lots of SWF use cases (rightfully so as a successor), but also more cases like service orchestration, etc.
Good to know it generalized, but seems it's only generalized on board games like problems where you have a problem space to search through, is that understanding correct? If so, it probably won't help in the use cases we are talking about here, right?
I didn't mean it's a requirement, but a fact that what ever DL could do is based on its input (labeled data), and that input has to come from human, or the task underneath is possible for human.
The reason I relate human skill with DL, is because for tasks where there are real patterns that humans can consistently recognize, DL is applicable, because it relies on human labeled inputs. For patterns, even humans can't consistently recognize (or if there are patterns at all), there's no way for DL to be better than humans, simply because the inputs from human can't be trusted. The example of DL does better job at e.g. recognizing dog breeds, recognizing cancers, whatever, is because those jobs are possible for humans to begin with. The only difference there would be DL could work on more human labeled data than any real human, so that they can recognize more patterns. But if those human labeled data/patterns dont exist in the first place, why we would believe DL could do a better job?
By DL, i meant generic DL techniques like CNN, LSTM (used in this article), which relies on large amount of labeled data to train, and predict on similar data, thus my comment about automation. Alpha Go (Zero) is very specialized for the game of go, not sure how much of its specialized algorithm could transfer to other generic use cases.
Total noob question in this space (algotrading with DL algorithms).
What DL is good at, is to automate tasks that are easy for humans (e.g. telling cat from dog, understand a sentence from sound wave, translation, etc) but hard for machines without DL.
Now, if a task is even hard/borderline impossible for humans to achieve (e.g. predicting stock market with higher accuracy and consistency), why would we believe DL could do a better job than humans?
Honest Question: is there enough data to show how likely the battery will burst into flames in crashes? So far the few incidences that I'm aware of make it feels like "the battery don't always burn, but when it does, there's no way to survive out of that". How true is that?
Qualcomm has been facing multi-front battles in the past few years: Apple Lawsuit, NXP deal nixed, Broadcom hostile take over, China problem, Regulator problems, Board In-fights, pressure from activist investor to "cut cost"/split the licensing business... etc, etc. Making the arm server business from 0 to a commercial hit would be too much to ask. Although it did have the best chance.
There's a documentary: The Science of Fasting ( https://www.youtube.com/watch?v=t1b08X-GvRs) which talks about fascinating effects of fasting therapy, and interesting effects of fasting on cancer treatment. Looks like it's inline with what's discussed here and in the comment section. On the flip side, I feel like fasting has the potential of becoming the new "meditation", but with more potential danger than just an over-hyped hip practice.
Not sure why you are downvoted, reading the 2nd paragraph you cited (and some other parts in the article), it's definitely hard to view them as comments only against the government instead of generalizing over a people. I guess since this post is mostly a rant (?), we could afford a less coherent statements. But it does feel a bit like one of those "not being a racist, but..." moments.
I see lots of evaluation of GCP, or comparison between GCP vs AWS recently. But I don't see as many for the Azure platform, although Azure being the 2nd in the cloud provider market, and even closing on AWS in market share. Anybody has any insights on this?
I guess the premise of the parent comment is that China, at least in the tech sector, is considered as "developed" as the US. If that assumption holds true, what the parent comment isn't far off. But that's an assumption that may or may not be true, depending on who you ask at least.
Sure, CDMA vs OFDM, definitely they can be regarded as "material difference" But that's not the logic/argument the original article made. They only used bandwidth and delay to quantify "material difference". Following their logic, they can make the same argument for 3G vs 4G back in the day.
The coverage/range of 5G depends on the scale of deployment, and the environment of the deployment (e.g. urban with lots of tall buildings, vs rural), assuming we are talking about mmw type of 5G not LTE+ type of 5G.
Due to higher frequency, 5G might actually have a disadvantage in range for a single tower, in dense urban environment due to weaker penetration. But it could be compensated by deploying more towers, at more strategic locations. The value proposition is that the gain in speed and delay would be well worth it.
Weak penetration is still a problem for long range open field 5G communication, but due to different factors like fog, weather conditions, etc.
Regarding data caps, I view it as more of a business model question, less related to the technical capabilities of 4G vs 5G.
When "4G" was initially commercialized, one could have claimed it has No "Material Difference Between 3G and 4G". HSPA(+) was just faster 3G, already reaching lower end of 4G bandwidth requirement, minimally qualifiable as 4G. But has 4G/LTE eventually turned out to be "one generation" better than 3G? Hell yeah.
Same thing here for 4G-5G transition. The real deal is when milimeter wave got massively deployed and commercialized.
Very interesting post. Could you elaborate more (maybe another blog?) on the DevOps perspectives on how to manage a prod Elixir/BEAM stack (CI/CD, scaling, dev vs test vs prod environment management)?
Definitely the things Linkedin do are so much worse and creepier, just by looking at the people they recommend me to connect to, the level of offline tracking they must be doing are creepy beyond imagination.
ah, thanks, that's what i wanted to ask. Since it's MQTT, how does it support iOS when the app is in the background? or it won't support that scenario on iOS? On Android, does it also use MQTT to maintain a connection? instead of using FCM/GCM? On the browser, how's the MQTT connection maintained? does it have service worker support for "offline" use cases?
Thanks for the response. I should have been more specific, what's AppSync's mechanism for push notification for the 3 platforms: FCM (GCM?) for Android? APNs for iOS? HTML5 Push API for browsers?
Facebook is indeed in a crisis, but I just got the feeling that the media is exaggerating anything that has the slightest negative bend to pour more fuel to the fire.