LLM
/llm2058
A space to discuss large language models, AI agents, and how they could interact with Farcaster data
Comparison tool for AI/LLM assisted search I got (mostly) working this weekend.
188955
Kyle
@banta·2 days ago
Decently productive weekend. Finally got a chunk of hours yesterday to hop in the zone and work on this project for my gf (is an SEO) to be ready for the rise in AI assisted search. Still a bit buggy and kinda slow, but felt good to hack around on a side project again!
Compare-bot.com
Compare-bot.com
https://www.twitch.tv/gemini_plays_pokemon
@rileybeans shared this link (via kimiya mag!! see link in reply to learn more) and all i can think of is @keccers.eth becoming even more mega brain
https://x.com/iliane_5/status/1906409499637813702?s=46
https://x.com/iliane_5/status/1906409499637813702?s=46
Heard from a guy who works in the AI agent space today that Microsoft is burning 25x of their Copilot revenue on inference costs currently in an attempt to catch up to Cursor.
Alright GTP Cooked with this update
243361
Apurv
@apurvkaushal·20:21 25/03/2025
Today I learnt #6
Myth: Text to image/ video tools like Midjourney, Sora are also built on large language models similar to GPT (next text prediction)
Fact: Text to image/ video is possible because of an algorithm class "diffusion models".
Myth: Text to image/ video tools like Midjourney, Sora are also built on large language models similar to GPT (next text prediction)
Fact: Text to image/ video is possible because of an algorithm class "diffusion models".
1/ Humanity’s bullshit detectors have gotten incredibly finely-tuned. We’re cutting information to the bone.
Example: LLMs convey information with far fewer words than other methods. One way of looking at this is they "gloss over" important subtext. Another way... 🧵👇
Example: LLMs convey information with far fewer words than other methods. One way of looking at this is they "gloss over" important subtext. Another way... 🧵👇
Does anybody have suggestions for a free Whisper LM app? My free trial on whisper flow is about to end.
Only AIs with memory across conversations can serve as long-term strategy consultants or conversational companions.
This will be a key differentiator for user retention.
This will be a key differentiator for user retention.
✨ Critical *Creative* Reasoning ✨
The limits of LLM’s when facing compositional tasks
https://www.quantamagazine.org/chatbot-software-begins-to-face-fundamental-limitations-20250131/
https://www.quantamagazine.org/chatbot-software-begins-to-face-fundamental-limitations-20250131/
Stop overpaying for AI responses! Route queries to the right model & cut costs.
By analyzing prompt complexity, an LLM router/gateway smartly chooses the most efficient model for each query. Essentially only serving more capable/expensive models when needed.
Here are a few open-source tools to get started!
By analyzing prompt complexity, an LLM router/gateway smartly chooses the most efficient model for each query. Essentially only serving more capable/expensive models when needed.
Here are a few open-source tools to get started!
Old DIY = Do It Yourself
New DIY = Delegate It Yourself
New DIY = Delegate It Yourself
what is missing in current virtuals agents ?
The most vulnerable jobs set to be disrupted by AI are the current set of AI's
chatgpt o1 operator's chrome tab chaos looks a lot like mine on any given day, get it together operator
Deepseek wins hands down
a post i wrote this week. curious for your thoughts https://silv.blog/future-of-529-plans-paying-for-ai-tutors-micro-credentials/
if you like deepseek r1 and want to talk to an uncensored version of it try:
https://huggingface.co/NaniDAO/deepseek-r1-qwen-2.5-32B-ablated
https://huggingface.co/NaniDAO/deepseek-r1-qwen-2.5-32B-ablated
☝🏻OpenAI has started offering free tokens in exchange for user data
A free limit of one million tokens per day, including the o1 model, is available on the OpenAI platform until the end of February.
However, to take advantage of this offer, users need to agree to have their prompts and answers used to train the company's AI models.
A free limit of one million tokens per day, including the o1 model, is available on the OpenAI platform until the end of February.
However, to take advantage of this offer, users need to agree to have their prompts and answers used to train the company's AI models.
found an interesting article on how to get the best summaries through chatGPT
The prompt is too large to share in 1 cast so i'll simply share the link to the arxiv article, you can find the prompt in page 2.
(arxiv link in the first reply)
The prompt is too large to share in 1 cast so i'll simply share the link to the arxiv article, you can find the prompt in page 2.
(arxiv link in the first reply)
a favorite workflow when i have a lot to untangle for work (i.e. high context tokens)
1. verbal diarrhea into voice memo app (on either macOS or iOS)
2. if I am worried that voice memo might crash, i might turn on fathom too and have a meeting by myself as a backup
3. copy transcript into notebooklm.google and query
1. verbal diarrhea into voice memo app (on either macOS or iOS)
2. if I am worried that voice memo might crash, i might turn on fathom too and have a meeting by myself as a backup
3. copy transcript into notebooklm.google and query
»#AI #Medical #NoteTaking Apps Enjoy Healthy Wave of Investment.« https://www.pymnts.com/healthcare/2025/ai-medical-note-taking-apps-enjoy-healthy-wave-of-investment/?eicker.news #tech #media
How to compile website content for LLMs study?
1️⃣ llmstxt - will compile any website into a single text file that can be passed to any LLM. https://llmstxt.firecrawl.dev/
2️⃣ Or get the copied text of any web page by entering https://r.jina.ai at the beginning of the web link
These two methods will help you with web content. Instead of complex and cluttered HTML code, it takes only the necessary text from the site, removing everything unnecessary: markup, scripts and styles, which simplifies AI training, improves the quality of answers.
1️⃣ llmstxt - will compile any website into a single text file that can be passed to any LLM. https://llmstxt.firecrawl.dev/
2️⃣ Or get the copied text of any web page by entering https://r.jina.ai at the beginning of the web link
These two methods will help you with web content. Instead of complex and cluttered HTML code, it takes only the necessary text from the site, removing everything unnecessary: markup, scripts and styles, which simplifies AI training, improves the quality of answers.
