rkwz

Strengths of open-weight models:

* Privacy: Can use sensitive data, can be used in companies/industries that have strict regulations

* Static model: Can "pin" the model version. Don't need to worry about underlying model being changed but having the same name

Not the parent, but I started using Llama 3.1 8b and it's very good.

I'd say it's as good as or better than GPT 3.5 based on my usage. Some benchmarks: https://ai.meta.com/blog/meta-llama-3-1/

Looking forward to try other models like Qwen and Phi in near future.

I think this would resonate with a lot of people, do you have a write up on what works and what doesn’t?

Genuine question, if they have so much excess capacity and low domestic demand, wouldn't it be easier to just export to other countries? Or are there barriers put in place in the world stage that limits this?

Data still needs to be cleaned ;)

ML based linkage has evolved a lot over the past years, and a few people are trying out LLM based ones - https://arxiv.org/pdf/2403.06434v1

TIL about WebTransport https://rxdb.info/articles/websockets-sse-polling-webrtc-web...

What are you intending to use it for?

> With such a huge organization, you had to go through 15 different stakeholders to get a single thing done, and there was an ingrained middle management whose only function was to connect you to the right person.

If you're able to recall, can you list who these 15 stakeholders were? Are they like legal department approval, etc?

> But, I want to build beautiful web applications, but I don't get CSS. Like is it supposed to be this hard to build a UI?

IMO, CSS is the hardest part of FE dev. It's deceptively easy, but you can still create good looking UI if you use 3rd party libs like bootstrap, tailwind, bulma etc.

> Recently, I've been trying to learn more about Frontend and started exploring a React project at work.

> Given my experience with this archaic pattern, I've now started exploring HTMX with Golang. This seems more my speed.

Do you want to build FE apps or learn about how FE frameworks? If you just want to build, you can use Golang templates with the above mentioned CSS libs. If you want to learn for career change (BE->FE), you'd have to learn React/Vue/Svelte etc.

> It is included in a "kettle body" kit which costs 80% of the price of a new kettle and is only sold to professional repairers.

Sounds more like malicious compliance

There are some benefits for people outside the field to propose ideas: It helps you think more thoroughly about why you're doing something certain way. Whether by forcing you to question your (blind) beliefs or making you find a simpler solution to the problem.

> find something popular and offer an incredibly mediocre alternative to it that will still be used by those teams who want to move as much as possible to AWS.

Any insights on _why_ they turned out to be mediocre? Is it because of them being released as MVP and staying that way for many years or just a lack of interest afterwards?

> Submenu is optional

> Apply to multiple types (files or directories; can be as granular as file extension)

> Show output (sometimes you want to see the results of the command)

> Confirmation before running (if it’s a potentially dangerous operation)

Nit: The last 2 points are actually possible in Automator

Some examples:

* Remove PDF password: https://www.sheshbabu.com/posts/remove-pdf-passwords-using-q...

* Summarize text with Ollama: https://www.sheshbabu.com/posts/system-wide-text-summarizati...

> GPT-3 (babbage-002)

I'm surprised babbage is still available via APIs - https://platform.openai.com/docs/models/gpt-base

Anyone else using this?

> Don't get me wrong, search has become extremely problematic, ... but how much effort does it take really?

Many days or weeks infact

The provided examples don't illustrate the problem properly. Let's say you're new to a problem domain but don't know it's a highly specialized domain. So you don't know the keywords to search but have a vague sense of what needs to be accomplished.

If the new tool can make sense of your vague problem and point you in right direction, it would be great.

Don't know if this is actually a $600B sized problem.

> I can’t do without the help of AI in my daily development work. I use ChatGPT, GitHub Copilot and various AI-driven design platforms. AI can help me improve my efficiency by at least 30%. For some HTML pages or pure algorithm modules, AI code accounts for 80%, and only a small amount of modification is required.

If you replace "AI" with "StackOverflow", would you still feel the same?

From the tweet: > Hey @AnthropicAI: I get you're hungry for data. Claude is really smart! But do you really need to hit our servers a million times in 24 hours? You're not only taking our content without paying, you're tying up our devops resources. Not cool.

Assuming they're literally getting hit "a million times in 24 hours", wouldn't it still cost bandwidth? Also, perhaps iFixit doesn't have such traffic in their past so they didn't bother optimzing prematurely.

I feel the blame mostly falls on Anthropic not being considerate when they scrape other people's content.

I've used Passport SAML for Node - https://github.com/node-saml/passport-saml

Maybe they're summarizing/processing the documents in a specific format instead of chatting? If they needed chat, might be easier to build using RAG?

Thanks, any ideas why it's not possible to build a generic eval for this? Since it's about asking a set of questions that's not public knowledge (or making stuff up) and check if the model says "I don't know"?

Great points!

> Not having retros for engineering processes (e.g. does our 80% coverage rule still make sense, should Jenkins be our job runner, etc.)

How is this usually done? Is this suggested by the team and get prioritized based on popularity or some other criteria?

> Trying to solve problems by having engineers "do better", rather than designing and implementing better processes

Can you elaborate on this?

> Longer context window (1M+)

What's your use case for this? Uploading multiple documents/books?

Perhaps if we think of LLMs as search engines (Google, Bing etc) then there's more money to be made by being the top generic search engine than the top specialized one (code search, papers search etc)

> A significant effort was also devoted to enhancing the model’s reasoning capabilities. One of the key focus areas during training was to minimize the model’s tendency to “hallucinate” or generate plausible-sounding but factually incorrect or irrelevant information. This was achieved by fine-tuning the model to be more cautious and discerning in its responses, ensuring that it provides reliable and accurate outputs.

Is there a benchmark or something similar that compares this "quality" across different models?

> Some kind of academic CRM where you can start a new 'project' with some keywords and it assembles highly cited works and the authors of the works. From there you can recursively search through the bibliography of the seminal works and the works that cite the seminal work to build a research map. When researching different fields you often end up finding a) who the top researchers are and then you want to go read all of their stuff. b) who is currently working on the thing in $current_year who you might want to contact and talk to.

Yes, this would be great. Would also prefer if there's a way to trace back to the "origin" papers from which other papers build upon, and if you don't understand a term or concept, can search papers that explain it better.

Email clients are still stuck in the 90s the way they render HTML, good luck asking them to render Markdown ;)

I use Merriweather as I feel it looks legible

How it looks like: https://www.sheshbabu.com

> I don't get these "privacy" based tracking / analytics systems, it sounds like an oxymoron to me.

Not really oxymoron if they don't track your PII like IP address, email etc. Don't know how Plausible works, but usually these systems track page views, etc without the associated visitors' data. They either anonymize the visitor metadata or just aggregate the site data like what urls are visited, etc.

If you have Mac and Ollama, you can use AppleScript to add summarization to any application - https://www.sheshbabu.com/posts/system-wide-text-summarizati...

Link to original post - https://www.reddit.com/r/Btechtards/comments/1axnhdz/deleted...

> On Mar 18 and 19, he ran a programmed script to delete 180 virtual servers in the system.

> The system that Kandula’s former team was managing was used to test new software and programs before launch. In a statement to CNA on Wednesday, NCS said it was a "standalone test system".

> As a result of his actions, NCS suffered a loss of S$917,832.

Wondering if these are CI/CD pipelines, and how the loss amount was calculated since these can be spun up again.