Follow-up post to https://hexbear.net/post/3852906

...or at least the kind of model that isn't fueled by burning a small forest for every query? I am wanting to play an old video game, but I'm still only just learning the language and could use an aid. I really, really want to avoid any of this incredibly wasteful AI stuff.

...or at least the kind of model that isn't fueled by burning a small forest for every query? I am wanting to play an old video game, but I'm still only just learning the language and could use an aid. I really, really want to avoid any of this incredibly wasteful AI stuff.

It feels like the US isn’t releasing what it has. I don’t think they’re behind, maybe just holding back?

It wasn't a health concern. The only concerns for 5G were long-term practicality, energy and service costs and shorter coverage.

5G was supposed to be made for robots, IoT and wireless home internet.

I haven't tried 5G, but it seems that the expected speeds would be the same as 4G in early days.

AI, as it currently exists, is terrible, devoid of much real value or use, and probably overall degrading our society (and definitely our climate). This is unlikely to change as the technology requires massive amounts of capital to train and run the models. Massive data centers that are an endless pit of energy, ghouls that want to use the technology for nefarious purposes such as extracting more wealth from you and I, to the extreme of facial recognition of "terrorist" suspects for drone strikes.

There is massive amounts of capital in this area, and other than the energy and equipment for running the models, another big cost for firms that develop LLMs is the need for training data that requires hundreds to thousands of people to generate manually. This is where you can potentially earn some decent remote money. There are caveats, and I by no means recommend this to anyone as a sole source of income. Work availability fluctuates week by week, the biggest company in this space is very evil, and you might end up spending a lot of time for very little benefit.

Your actual ability to earn depends mostly on your country of residence your level of education. In the USA, with a M.S. in Chemistry (they have never confirmed any of my education btw), I have been able to earn at most $60/hr, but more often $25-35/hr. The pay changes constantly because every project pays differently, and you will change projects very often, usually every week or two. I have also gone through a 2 month dry spell of having no/little work to do, which is why this is not something I would ever recommend as a source of income you rely on.

What is AI training and why should I care?

Specifically, AI training encompasses any work that involves creation or annotation of data that is fed back into an LLM model to improve it's capabilities. The data can be images, audio, video, and most commonly text. The specific task at hand can vary widely, but the most common involves rating various dimensions of the LLM's response and then improving the response.

How exactly is this used by a company? As a simple example, imagine that you are creating a service that uses an LLM to create recipes from a list of ingredients. First, you might download every cookbook and recipe that you can find by crawling the web, and feed that into the model. Unfortunately, the quality of the recipes might not be very well standardized, some transcription errors may have occurred, or you might have other specifications that you would like your model to follow. So you send those recipes, one-by-one, to thousands of people who will fix any errors, add additional context, and make sure that the recipes fit the specifications. Now this data set can be used to fine-tune the model and hopefully improve it by some measure.

Companies I have worked for

Scale Labs (Remotasks, Outlier)

Scale Labs is by far the largest company in this space, and is, at least up until recently, the de facto monopoly. It was founded by Alexandr Wang in 2016, after he dropped out from MIT at the age of 19. The company grew by setting up large centers (aka digital sweatshops) in Sub-Saharan Africa and Southeast Asia, paying the lowest possible wage for workers to annotate images from autonomous vehicles. Imagine sitting in a sweltering computer center, getting paid well below $1 an hour, to circle cones, pedestrians, vehicles, bikes, over and over again, with no job security - that is how this little psychotic dipshit become a billionaire at the age of 24. There are numerous articles out there about how bad this company is, and of course they are also the official AI partner of the DoD.

Scale runs two tasking platforms - Remotasks, and Outlier, the latter of which myself and @corgiwithalaptop have used. They seem focused on recruiting people with specific expertise, for a list of all fields you can look at their site (outlier.ai). Although I was hired as a "Chemistry Expert", most of my projects haven't involved Chemistry at all.

There are some major problems with this platform:

• You will have no tasks (empty queue, or EQ) for extended periods of time (sometimes weeks to months)
• "Training" is usually unpaid and extreme bullshit
  • You need to do new training for each new project. The training documents are often vague or contradictory. You have to do graded quizzes and graded assessment tasks that you will fail, and never get feedback on why you failed. This is a very common experience, I once had a single week where I failed 4 assessments in a row, meaning I did 4 unpaid trainings for no reward.
• The platform itself is full of bugs and technical issues
  • Once I was assigned to a project, but a bug prevented tasks from actually being allocated to me. The project's managers (queue managers, or QMs) told me that it is a bug that I would have to ask support about. By the time support got back to me over a week later, the project had wrapped up. I got assigned to a new project, and faced the exact same bug, which took another week to get fixed. This entire time, I am unable to do any work on the platform or make any money.
  • If there is a technical issue that prevents you from completing a task, you will not get paid at all for the time you had already invested in the task.
• Every task gets reviewed and you are given a rating out of 5. Reviewers are just normal taskers that get promoted to reviewer willy-nilly (I have been a reviewer on many projects), and some reviewers are just god awful. They will straight up not understand your prompt, or not be familiar with some concept in the task, and give you a 1/5 for no reason. I have even been on projects where people were using an LLM to automatically complete the reviews, giving erroneous bad scores. If you get enough low scores, you can be removed from the project.
• Support is non-existent and useless. It takes days to weeks to get any response, and sometimes the response is just a canned response where you can tell they didn't actually read your ticket.
• They can remove you from the platform at any time for any reason, and anecdotally I have heard about this happening to people who did nothing wrong. There is no recourse if this happens to you, you can reach out to support but like I said, this is useless

Despite these problems, I am currently making money on the platform and will continue to do so as long as I can. For all of these reasons, I can not recommend outlier unless you have infinite spare time and really want a remote way to make money. If you have an expertise, especially coding/math, I think you might have a better chance of getting in and making good money. Recently, they have been promising to make changes to address these issues, but scale labs is ultimately an authoritarian employer that has no incentive to make their workers lives better, and only has incentive to increase the efficiency of labor extraction, so take that with a grain of salt. If you need income, I would not recommend wasting to much time on it or expecting anything out of it, but it might be worth a try.

Stellar AI

I have only been on this platform for about a week, and am only on one project so far, so I don't have much to say other then the following ways it is better than Scale's platforms:

• Training was much better. Instead of just reading a google doc, they walk you through exactly how to do a task within the tasking interface. It was also paid
• If there is a technical issue that prevents you from completing a task, they will still pay you for your time My current project involves guiding the AI through navigating websites to get information to answer a prompt, for example, "what is the cheapest airline to fly to Hawaii," it is not all that difficult and the pay is $25/hr.

Conclusions

If you are good at writing-based work, and want remote ways to make money, and have plenty of spare time, it might be worth trying out these platforms. The worst case scenario is that you spend some time on it and ultimately reap no benefit, but there is a chance you get in and can successfully make decent money doing work that really isn't that difficult. The majority of the stress I have had is due to the Outlier platform itself.

If you have special expertise (any advanced degree, multiple languages, especially if you have coding experience) you might have better chances. Pay rates vary by country, so if you aren't based in the USA I am not sure what the earnings potential is.

@corgiwithalaptop also works for Outlier and they might be able to offer a different perspective. If you do decide you want to apply for Outlier, reach out to me or them and we can give a referral code - if you use my code and get the job and complete 10 hrs of work I would get $200, which I pledge to give $100 through the mutual aid comm here and the other $100 to my local mutual aid.

Test-time training (TTT) significantly enhances language models' abstract reasoning, improving accuracy up to 6x on the Abstraction and Reasoning Corpus (ARC). Key factors for successful TTT include initial fine-tuning, auxiliary tasks, and per-instance training. Applying TTT to an 8B-parameter model boosts accuracy to 53% on ARC's public validation set, nearly 25% better than previous public, neural approaches. Ensemble with recent program generation methods achieves 61.9% accuracy, matching average human scores. This suggests that, in addition to explicit symbolic search, test-time training on few-shot examples significantly improves abstract reasoning in neural language models.

Budget Chinese retailers such as Temu and Shein successfully made Amazon feel insecure enough for it to come up with its own version, Amazon Haul. The e-commerce giant had been talking about launching a similar discount store with all items priced under $20. It launched in beta today and is available as a separate tab on the Amazon phone app, titled Haul.

It kept its promise and ensured a $20 price cap with most items under $10. The product categories include fashion, home, lifestyle, and electronics. It says the delivery time will be between one and two weeks, much longer than Amazon’s typical delivery timelines of within a week and a day or two if you’re a Prime member. However, the longer delivery times make sense since the Amazon Haul products will be shipped directly from a fulfillment center in Guangdong, China.

Scrolling through the items, you’ll quickly realize they’re fairly poor quality but nothing different from what Temu and Shein offer. Unlike the regular Amazon, Amazon Haul is more for impulsive shopping during late-night doom scrolling. This might be why the feature is mobile-exclusive. It’s similar to the in-app shopping feature TikTok introduced because it knows how impulse shopping works.

Unlike traditional language models that only learn from textual data, ESM3 learns from discrete tokens representing the sequence, three-dimensional structure, and biological function of proteins. The model views proteins as existing in an organized space where each protein is adjacent to every other protein that differs by a single mutation event.

They used it to "evolve" a novel protein that acts similarly to others found in nature, while being structurally unique.

America spends so much more money than anyone else in every area of production and gets so little for it. for sure not corruption though as we have carefully, narrowly defined it

I have various ebooks and audiobooks purchased from a certain monopoly that I’ve stripped drm from. If I were to share them, I’d first want to check & strip the metadata to ensure there are no identifiers in there. Any suggestions on how to do this?

What got me thinking about this with PoC||gtfo article about metadata

STOP TRYING TO INSTALL COPILOT

Elon Musk’s support for Donald Trump is set to boost X’s flagging business, with some marketers poised for a return to the social media platform in order to seek favor with the incoming administration.

Media executives told the Financial Times that some brands were preparing to advertise on X once again, as its billionaire owner was likely to gain an influential role within a second Trump White House.

Lou Paskalis, chief executive of marketing consultancy AJL Advisory and a former media executive at Bank of America, said some marketers are likely to reallocate spending back to X as “political leverage,” such as if they were seeking government contracts. He added companies would seek to get in the “good graces of Elon,” who has been given a wide remit by Trump as co-head of a new Department of Government Efficiency.

“It could be seen as an official channel for White House communications,” said another advertising agency chief, adding that Trump’s victory has lent Musk new legitimacy as well as power over brands in sectors that could face new regulatory curbs from Trump.

Anyone have experience with it? Is it okay?

Update: I will not be joining bluesky lol

Ask the search engine Ecosia about “Paris to Prague” and flight booking websites dominate the results. Ecosia’s CEO, Christian Kroll, would prefer to present more train options, which he considers better for the environment. But because its results are licensed from Google and Microsoft’s Bing, Ecosia has little control over what’s shown. Kroll is ready for that to change.

The Berlin-based company, which donates its profits to tree planting, and its Paris-based competitor Qwant are announcing Tuesday that they will team up to develop an index of the web.

The for-profit joint venture, dubbed European Search Perspective and located in Paris, could allow the small companies and any others that decide to join up to reduce their reliance on Google and Bing and serve results that are better tailored to their companies’ missions and Europeans’ tastes. “We could derank results from unethical or unsustainable companies and rank good companies higher,” Kroll says of the eco-minded Ecosia.

Dems using their lame duck session in the Senate to pass some slop for their donors.

cross-posted from: https://hexbear.net/post/3884192

For the purpose of miniature wargaming, I’m considering getting into 3D printing because there are tons of amazing sculpts I’d love to get my hands on.

A couple of questions, though

I’ll admit I know next to nothing about 3D printing, but I could spend up to 300 eurodollars (or a bit more) on a 3D printer. Are there any key things to consider when picking up a printer? I don’t need the cheapest model, as price isn’t the main reason I’m getting into 3D printing. I care quite a bit about the quality of the finished models.

I live in an apartment, so does this even work logistically? I’m aware there are some health hazards with resin/3D printing—how serious should my concerns be about that? (That’s why I’m asking here and not in a 3D printing subreddit, where folks might be biased.)

Are there any recommended communities or YouTube channels for 3D printing? As I mentioned, I’ll be using this primarily for gaming miniatures and possibly wargaming terrain, not for other 3D-printable items.

Edit: I found it for $36 elsewhere. :D