Garcia98

As a Spaniard who was about to take a well-deserved siesta, I found these points so blatantly inaccurate that I had to get up and address them.

>3% tax on my global wealth every year (in effect it's more like 4.5% because there is also significant capital gain tax)

This only affects multi-millionaires, and even then, your numbers are wrong. The national wealth tax only applies to net worth over €3M. The top rate of 3.5% is only for assets over €10.7M. And some regions like Madrid and Andalusia offer a 100% exemption from the regional tax.

>60% tax on a semi-decent tech salary

False. The top income tax bracket is ~47% on earnings over €300K, which is far beyond a "semi-decent" salary here. More importantly, as a foreigner, you can use the "Beckham Law" to pay a flat tax of 24% on your first €600K of income for six years.

>25%+ capital gain tax for my investors (if I find any)

You're cherry-picking the highest rate again. It's a progressive tax starting at 19%. It only exceeds 25% on gains over €200K. Besides, if your investors are not in Spain, they are taxed in their country of residence under its tax treaties with Spain.

>people don't speak English, many of them are proud of it

We are not French, we won't scold you for trying to speak English. While not everyone is fluent in rural areas, people will genuinely try their best to help you, even if it's through Google Translate.

>scheduling anything with anybody is impossible, the concept of being on time doesn't exist

You're confusing social life with professional life. Yes, we might be 15 minutes late for a beer, but in a business context, being late is just as unprofessional here as it is in London or New York.

>Bureaucracy is slow, full of paperwork, can't be done online, can't be done in English, the result depends on clerk's mood on a given day

Spanish bureaucracy is as slow as in anywhere else, but it is indeed digitized. You are required to file your tax forms (available in English btw) digitally, and you can register a business, get your SSN, etc., with a digital certificate from your computer. I've only had to show up in person once in the last five years to register my address after moving to a new city.

>if someone comes to my house when I am away I can't kick them out, need to find another house and keep paying bills for the new occupa(s)nts

This is just rage bait. What you are describing, someone entering your primary home, is trespassing, a criminal offense. The police will have them evicted within 48 hours. The infamous "okupa" issue applies to properties that are clearly abandoned or second homes left empty for years, not your actual residence.

Pretraining on the Test Set Is All You Need

https://arxiv.org/abs/2309.08632

The main problem with the non-Apple laptop market is that there is a mind-boggling number of confusing models, SKUs, processor/gpu variants, etc., and wildly variable physical quality control that confuse consumers and leave them unhappy. This is the flip side of choice in prioritizing, say, gaming performance over battery life while optimizing price or vice-versa.

This is 100% it, Lenovo has been killing it lately with their Yoga/Slim series, but for every laptop they have that competes with a MacBook, they also have a myriad of other options that are just e-waste. At the end of the day, the average consumer is not going to do the same kind of research that a tech enthusiast might do, and Apple has a somewhat simple catalog (although incredibly overpriced once you step out of the entry configs).

I see this (and other ublue images) as an alternative to Ansible, rather than just an image. I could fork this repo, automate PRs from upstream with a GH action while making my own changes to it and keep an automated CI/CD pipeline. This painless extensibility is a big advantage imo over traditional distributions.

Have you ever tried mamba[0]? It's much, much faster than conda for me.

0: https://github.com/mamba-org/mamba

The issue is not AI, nor browser extensions per se, the issue is the lackluster permission system that Chrome extensions have, it's pretty similar to what Android had 7 (?) years ago, which should not be acceptable in 2023.

You can also submit this Google Forms buried three links deep in their privacy policy: https://docs.google.com/forms/d/e/1FAIpQLScrnC-_A7JFs4LbIuze...

Note that the FLAN-T5 variant in 10th place is 3B parameters large, there is another FLAN-T5 variant that has 11B parameters that should perform better (and the fact that a 3B model is able to compete with Alpaca-13B is impressive by itself).

It is the same in Spain.

Fun story, I had a professor in university that tried to force us to buy his own book for his course, refusing to provide notes or any other free alternatives, and for the final exam he allowed anyone to bring in his book.

We complained to the school board as this was unfair and there was a conflict of interests, the exam was redone by a committee and a new professor replaced him the next year.

I have no clue where the data for Spain comes from, in June 2022 the average price for the regulated market was 0.291347 €/kWh[0], that's ignoring the free market rates which are what 70% of Spaniards use (and they are cheaper when the market is volatile, like last year).

[0] https://www.ocu.org/vivienda-y-energia/gas-luz/informe/preci...

Spaniard here, on my last bill I paid 0.120268 €/kWh, admittedly I'm on the free market but even on the regulated market the average price for March was 0.158719 €/kWh.

Is electricity really that cheap in the US? This chart shows that the average electricity price is 0.165 $/kWh, so I'm not sure what I'm missing: https://www.bls.gov/regions/midwest/data/averageenergyprices...

Providing a checksum along with the binary singlehandedly solves your concerns, and you can add gpg signatures if you want.

The author is overly optimistic with the current state of open source LLMs, (e.g., Koala is very far away from matching ChatGPT performance). However, I agree with their spirit, Google has been one of the most important contributors to the development of LLMs and until recently they've been open sharing their model weights under permissive licenses, they should not backtrack to closed source.

OpenAI has a huge lead in the closed source ecosystem, Google's best bet is to take over the open source ecosystem and build on top of it, they are still not late. Llama based models don't have a permissive license, and a free model that is mildly superior to Llama could be game changing.

The most notable paper is probably the one that defined the concept of transformers, Attention Is All You Need: https://arxiv.org/abs/1706.03762

LaMini-Flan-T5-783M outperforming Alpaca-7B in human evaluation is impressive, and I wish Flan-T5 got some more love from the community, there's too much buzz around Llama based models that could be invested into improving a more open model (Flan-T5 is licensed under Apache 2.0 vs the restrictive license of Llama).

Now I wish someone fine-tuned Flan-T5 on the dataset from Open Assistant, that could be a truly open Alpaca competitor.

I use a GPU server and runtime is not free unfortunately.

5-shot prompting.

I disagree, they made the decision to use datasets with restrictive licensing, jumping the alpaca/gpt4all/sharegpt bandwagon.

They also chose to toot their horn about how open-source their models are, even though for practical uses half of their released models are not more open source than a leaked copy of LLaMa.

I would have compared it to the fine-tuned version if it had been released under a truly open-source license. I think developers implementing LLMs care more about licensing than about the underlying details of the model.

Also t5-base is 220M params vs 3B params of stablelm, not really a fair comparison anyways.

I really dislike this approach of announcing new models that some companies have taken, they don't mention evaluation results or performance of the model, but instead talk about how "transparent", "accessible" and "supportive" these models are.

Anyway, I have benchmarked stablelm-base-alpha-3b (the open-source version, not the fine-tuned one which is under a NC license) using the MMLU benchmark and the results are rather underwhelming compared to other open source models:

- stablelm-base-alpha-3b (3B params): 25.6% average accuracy

- flan-t5-xl (3B params): 49.3% average accuracy

- flan-t5-small (80M params): 29.4% average accuracy

MMLU is just one benchmark, but based on the blog post, I don't think it will yield much better results in others. I'll leave links to the MMLU results of other proprietary[0] and open-access[1] models (results may vary by ±2% depending on the parameters used during inference).

[0] https://paperswithcode.com/sota/multi-task-language-understa...

[1] https://github.com/declare-lab/flan-eval/blob/main/mmlu.py#L...

What's the most cost-effective alternative to Paperspace? I had a nightmarish experience with them last week after my account got locked up twice when I was training a model with a 1.5 GB dataset that somewhere contained the string "Minecraft Server".

Sure, but gddccontrol has a GUI, has been supported since 2004 and works fine in 2023, that's why I find it a bit odd that he didn't know what to recommend to Linux users until now.

gddccontrol has been around for ages on Linux and it's not exclusive to a single desktop environment (and I'm sure there's a plethora of alternatives, but that's just the one I use).

I've seen this question asked repeatedly in many LLaMa threads, currently the best models that are truly open are the released models from the Flan family by Google, which includes Flan-T5[0] and Flan-UL2[1]. According to its paper, Flan-UL2 performs slightly better than Flan-T5-XXL.

These models perform slightly better than GPT-3 under some tasks[2], but they're still far from achieving the results from GPT-3.5 and GPT-4. This becomes evident when you try to use them in the real world; they're not "good enough" for general use cases, unlike ChatGPT models. However, if you can restrict your use case to one particular domain, you can achieve pretty good results by further fine-tuning these models.

[0] https://huggingface.co/google/flan-t5-xxl

[1] https://huggingface.co/google/flan-ul2

[2] https://paperswithcode.com/sota/multi-task-language-understa...

It's all client side, it downloads a model and performs the inference using ONNX runtime (which uses WASM).

Sorry for the late reply, as I said Flan-UL2 (or Flan-T5 if you want lighter models) fine-tuned against a dataset like CodeAlpaca's[0] is probably the best solution if it's intended for commercial use (otherwise LLaMa should perform better).

[0] https://github.com/sahil280114/codealpaca

Google's Flan-T5, Flan-UL2 and derivatives, are so far the most promising open (including commercial use) models that I have tried, however they are very "general purpose" and don't perform well in specific tasks like code understanding or generation. You could fine-tune Flan-T5 with a dataset that suits your specific task and get much better results, as shown by Flan-Alpaca.

Sadly, there's no open model yet that acts like a Swiss knife and gets good-enough results for multiple use cases.

I've been following open source LLMs for a while and at first glance this doesn't seem too powerful compared to other open models, Flan-Alpaca[0] is licensed under Apache 2.0, and it seems to perform much better. Although I'm not sure about the legalities about that licensing, since it's basically Flan-T5 fine-tuned using the Alpaca dataset (which is under a Non-Commercial license).

Nonetheless, it's exciting to see all these open models popping up, and I hope that a LLM equivalent to Stable Diffusion comes sooner than later.

[0] https://github.com/declare-lab/flan-alpaca

How does this compare to this project: https://github.com/visheratin/web-ai

Does it support using custom ONNX models?