<![CDATA[Sinisterly - Artificial Intelligence]]> https://sinister.li/ Fri, 25 Sep 2026 01:56:15 +0000 MyBB <![CDATA[Ai Project]]> https://sinister.li/Thread-Ai-Project Mon, 21 Sep 2026 17:11:09 +0000 Snickerdoodle]]> https://sinister.li/Thread-Ai-Project
I've been working on making an AI-powered app for Tiktok that grabs user's Tiktok videos and uses components similar to Creator Search Insight and general outputs from your typical AI (Chatgpt, Gemini, Grok) to redesign and create better optimized videos. The end goal would be to increase user's SEO and help users gain more likes, views, and visibility on their content regardless of the following.

I've already setup the database, the site and the connection between Tiktok developer app to the site as well.

I'm wondering what features should I asd rhat would make it more unique and more useful than the currently existing creator search insight or people's general access to AI.]]>
I've been working on making an AI-powered app for Tiktok that grabs user's Tiktok videos and uses components similar to Creator Search Insight and general outputs from your typical AI (Chatgpt, Gemini, Grok) to redesign and create better optimized videos. The end goal would be to increase user's SEO and help users gain more likes, views, and visibility on their content regardless of the following.

I've already setup the database, the site and the connection between Tiktok developer app to the site as well.

I'm wondering what features should I asd rhat would make it more unique and more useful than the currently existing creator search insight or people's general access to AI.]]> <![CDATA[AI Agent University]]> https://sinister.li/Thread-AI-Agent-University Mon, 21 Sep 2026 07:42:53 +0000 johnathon]]> https://sinister.li/Thread-AI-Agent-University
Also RIP Aaron Schwartz]]>
Also RIP Aaron Schwartz]]> <![CDATA[What has gotten you a job in AI?]]> https://sinister.li/Thread-What-has-gotten-you-a-job-in-AI Sun, 20 Sep 2026 21:08:11 +0000 Snickerdoodle]]> https://sinister.li/Thread-What-has-gotten-you-a-job-in-AI <![CDATA[How I've using AI to assist me in writing TTRPG Campaigns]]> https://sinister.li/Thread-How-I-ve-using-AI-to-assist-me-in-writing-TTRPG-Campaigns Thu, 17 Sep 2026 18:20:22 +0000 Skullmeat]]> https://sinister.li/Thread-How-I-ve-using-AI-to-assist-me-in-writing-TTRPG-Campaigns
I also have begun to use tools like stable diffusion to generate character profiles and locations, as a tool in the same vein, additions that add to the overall polish, never as a full replacement for a setting or character. Simply to help me flesh out what's already in my head.

There's some debate about the use of AI in TTRPGS, what do you think?]]>
I also have begun to use tools like stable diffusion to generate character profiles and locations, as a tool in the same vein, additions that add to the overall polish, never as a full replacement for a setting or character. Simply to help me flesh out what's already in my head.

There's some debate about the use of AI in TTRPGS, what do you think?]]> <![CDATA[AI is the new photoshop]]> https://sinister.li/Thread-AI-is-the-new-photoshop Mon, 14 Sep 2026 01:51:20 +0000 Snickerdoodle]]> https://sinister.li/Thread-AI-is-the-new-photoshop AI Is the New Photoshop

Photoshop changed image editing by letting anyone manipulate pictures without needing a darkroom.

AI is taking it to another level.

Instead of spending hours editing, you can just tell AI what you want. Remove something, change the background, fix a photo, create a new image, or completely change the style.

Photoshop gave us control over the image. AI gives us control over the idea. Sometimes a picture you take could have simple imperfections or a screenshot may be shotty, but apps like ChatGPT could easily mend those images. Even Stitch could merge multiple screenshots into a singular screenshot using overlapping frames and ai.]]> AI Is the New Photoshop

Photoshop changed image editing by letting anyone manipulate pictures without needing a darkroom.

AI is taking it to another level.

Instead of spending hours editing, you can just tell AI what you want. Remove something, change the background, fix a photo, create a new image, or completely change the style.

Photoshop gave us control over the image. AI gives us control over the idea. Sometimes a picture you take could have simple imperfections or a screenshot may be shotty, but apps like ChatGPT could easily mend those images. Even Stitch could merge multiple screenshots into a singular screenshot using overlapping frames and ai.]]> <![CDATA[AI Apocalypse still hasn't happened]]> https://sinister.li/Thread-AI-Apocalypse-still-hasn-t-happened Tue, 08 Sep 2026 04:02:09 +0000 Snickerdoodle]]> https://sinister.li/Thread-AI-Apocalypse-still-hasn-t-happened
The AI Job Apocalypse Still Hasn’t Happened

For years we’ve been hearing that AI was going to destroy huge numbers of jobs almost overnight. That hasn’t happened. The U.S. labor market is still relatively strong, and there hasn’t been widespread unemployment directly caused by AI. Even Sam Altman recently said AI hasn’t taken as many white collar jobs as he originally feared.

However, there's a catch. AI seems to be hitting younger workers and entry level positions much harder. In a recent Stanford study found employment for workers aged 22–25 in highly AI exposed jobs is about 19% lower than it would have been compared with less-exposed fields. So maybe the AI job apocalypse isn’t actually happening the way people predicted.

Instead of replacing millions of experienced workers overnight, AI could be making it harder for young people to get their first job in the first place. That could end up being a much slower but bigger problem.

Check out these articles I read for my sources peeps:

https://digitaleconomy.stanford.edu/news/canariesaug26/

https://www.newyorker.com/news/the-finan...-postponed

https://www.reuters.com/world/asia-pacif...026-05-26/
]]>
The AI Job Apocalypse Still Hasn’t Happened

For years we’ve been hearing that AI was going to destroy huge numbers of jobs almost overnight. That hasn’t happened. The U.S. labor market is still relatively strong, and there hasn’t been widespread unemployment directly caused by AI. Even Sam Altman recently said AI hasn’t taken as many white collar jobs as he originally feared.

However, there's a catch. AI seems to be hitting younger workers and entry level positions much harder. In a recent Stanford study found employment for workers aged 22–25 in highly AI exposed jobs is about 19% lower than it would have been compared with less-exposed fields. So maybe the AI job apocalypse isn’t actually happening the way people predicted.

Instead of replacing millions of experienced workers overnight, AI could be making it harder for young people to get their first job in the first place. That could end up being a much slower but bigger problem.

Check out these articles I read for my sources peeps:

https://digitaleconomy.stanford.edu/news/canariesaug26/

https://www.newyorker.com/news/the-finan...-postponed

https://www.reuters.com/world/asia-pacif...026-05-26/
]]> <![CDATA[Elon Musk on Artificial Intelligence - The Economist (July 2026)]]> https://sinister.li/Thread-Elon-Musk-on-Artificial-Intelligence-The-Economist-July-2026 Mon, 07 Sep 2026 19:12:09 +0000 Oni]]> https://sinister.li/Thread-Elon-Musk-on-Artificial-Intelligence-The-Economist-July-2026


Timestamps:
00:00 - AI will be smarter than humans in five years
12:10 - How AI companies could regulate themselves
15:20 - Will China be the leader in AI?
25:43 - Elon Musk on Sam Altman and the AI bosses
31:20 - “Work is going to be optional”
41:07 - Should one man have so much power?
49:44 - Starlink and the war in Ukraine
55:18 - “I got too involved in politics”
57:14 - Musk: zero people died because of USAID cuts
01:01:15 - Europe, the far right and civil war in Britain
01:08:21 - Musk: I’m not a racist
01:20:27 - What do people get most wrong about Musk?]]>


Timestamps:
00:00 - AI will be smarter than humans in five years
12:10 - How AI companies could regulate themselves
15:20 - Will China be the leader in AI?
25:43 - Elon Musk on Sam Altman and the AI bosses
31:20 - “Work is going to be optional”
41:07 - Should one man have so much power?
49:44 - Starlink and the war in Ukraine
55:18 - “I got too involved in politics”
57:14 - Musk: zero people died because of USAID cuts
01:01:15 - Europe, the far right and civil war in Britain
01:08:21 - Musk: I’m not a racist
01:20:27 - What do people get most wrong about Musk?]]> <![CDATA[Welcome to Artificial Intelligence]]> https://sinister.li/Thread-Tutorial-Welcome-to-Artificial-Intelligence Mon, 07 Sep 2026 19:02:12 +0000 Snickerdoodle]]> https://sinister.li/Thread-Tutorial-Welcome-to-Artificial-Intelligence
JUMBO WUMBO MEGA SUPER DUPER THREAD YALL

😭 on 😩

ARTIFICIAL INTELLIGENCE, LLMS & FRONTIER ARCHITECTURES


[Image: ai-you-mean.jpg]

PART 1: THE FOUNDATIONAL MECHANICS (WHAT IS AI vs. LLMs)


Artificial Intelligence (AI) is basically the bigger field of building machines that can do things we normally associate with human intelligence. That can mean recognizing images, understanding speech, making decisions, translating languages, and a lot more.

Large Language Models (LLMs) are one type of AI. They are mainly built to understand and generate language. Most of the major LLMs today are based on the Transformer architecture, which was introduced in 2017.

[Image: The-Transformer-model-architecture.png]

How LLMs Actually Work Under the Hood:

Tokenization: Before an LLM can work with text, it has to break that text down into smaller pieces called tokens. A token can be a whole word, part of a word, a number, punctuation, etc. Those tokens are then turned into numbers the model can process.

High-Dimensional Vector Embeddings: Those tokens are represented as vectors in a large mathematical space. This is one of the ways the model can learn relationships between words, concepts, and patterns.

Self-Attention Mechanism: This is one of the biggest ideas behind Transformers. Attention lets the model look at other tokens and figure out which ones matter most to the token it is currently processing. This is what helps it connect information across a sentence, paragraph, or much larger context.

Next-Token Prediction & Sampling: At the most basic level, an LLM predicts what token should come next based on everything that came before it. Settings like Temperature and Top-P can change how predictable or varied the generated answer is.


Recommended Video Breakdown:



PART 2: MODALITIES & ARCHITECTURAL VARIATIONS


AI isn’t just about typing into a chatbot anymore. Modern models can work with text, images, audio, video, code, and other types of information.

Multimodal Foundation Models: These models can work with more than one type of input. For example, a model might be able to understand text and images at the same time, or work with audio and video. Newer systems are increasingly designed around multiple modalities instead of treating every type of data as a completely separate problem.

Dense vs. Mixture-of-Experts (MoE):

Dense Models: A dense model generally uses the same set of parameters when processing each token. As the model gets bigger, this can make it more expensive to run.

MoE Models: MoE models split the network into different “experts.” A router network decides which experts should handle each token. This means the model can have a huge number of total parameters without having to activate all of them for every single token.

[Image: 01_moe_layer.png]

Reasoning / Test-Time Compute Models:

This is where things get interesting. Some newer AI systems use extra computing power while answering a question instead of immediately producing an answer. The model can spend more time working through a difficult problem before giving the final response.

This can be especially useful for things like math, coding, and complicated reasoning, although the tradeoff is more computation and usually more time.

Diffusion Models:

Diffusion models are behind a lot of modern AI image generation. The basic idea is that the model starts with noise and gradually turns that noise into a usable image or other output based on the instructions it was given.


[Image: 1769515884530?e=2147483647&v=beta&t=t7i3...EJ659OHqTQ]

PART 3: FRONTIER MODEL MATRIX & COMPARISON


Code:
+–––––––––––+———————————––+———————————–+––––––––––+ | Model Family | Core Strengths | Ideal Use-Cases | Deployment Style | +–––––––––––+———————————––+———————————–+––––––––––+ | OpenAI GPT Series | Tool use, multimodal capabilities, | General assistants, enterprise | Proprietary API | | | broad ecosystem | workflows, agent systems | | +–––––––––––+———————————––+———————————–+––––––––––+ | Anthropic Claude | Coding, long-context work, | Software engineering, analysis, | Proprietary API | | | detailed writing | research, drafting | | +–––––––––––+———————————––+———————————–+––––––––––+ | Google Gemini | Multimodal capabilities, large | Video/audio analysis, coding, | Proprietary API / | | | context, Google ecosystem | large-scale document analysis | Vertex AI | +–––––––––––+———————————––+———————————–+––––––––––+ | Open-Weight Models | Local deployment, customization, | Self-hosting, local development, | Self-Hosted / | | (DeepSeek/Meta | fine-tuning, cost control | private deployments | Local Execution | | Llama) | | | | +–––––––––––+———————————––+———————————–+––––––––––+


PART 4: PROMPT ENGINEERING & CONTROLLABILITY


A good prompt doesn’t have to be complicated. The main thing is telling the model what you want, giving it the information it needs, and making the expected format clear.

The C-R-E-A-T-E Prompt Framework:

C - Context: Give the model the background it needs to understand the situation.

R - Role: Tell it what type of expertise or perspective it should use.

E - Explicit Instructions: Say exactly what you want it to do and anything you don’t want it to do.

A - Audience: Tell it who the final answer is supposed to be for.

T - Template/Formatting: Tell it how you want the answer formatted. This could be a table, JSON, Markdown, code, etc.

E - Examples (Few-Shot Prompting): Give examples when you want the model to follow a particular style or output pattern.

Advanced Control Parameters:

Temperature: Controls how much randomness is used when selecting the next token. Lower values generally make the output more predictable, while higher values allow for more variation.

System Prompts: Higher-priority instructions that define how the model should behave, what rules it should follow, and what its overall job is.


[Image: 1767287671311?e=2147483647&v=beta&t=s6HT...BWlqmQFJm4]

PART 5: WHAT YOU DIDN’T THINK OF (THE REAL BLEEDING EDGE)


1. RAG (Retrieval-Augmented Generation)

One of the biggest problems with LLMs is that they don’t automatically know everything you need them to know. They can also make things up.

Retrieval-Augmented Generation (RAG) helps with this by connecting the model to an outside source of information, often through a retrieval system and vector database.

When you ask something:

Step 1: The system turns your question into a vector representation.

Step 2: It searches the available documents or knowledge base for information that matches the question.

Step 3: The relevant information is sent to the model as additional context.

[Image: RAG_schema.svg]

2. Agentic Frameworks & Function Calling

This is another major step forward. Instead of an AI only giving you text, you can connect it to actual tools.

Tool/Function Calling lets the model output structured instructions telling another system which function it wants to use and what information that function needs.

For example:

Code:
{“action”: “book_flight”, “date”: “2026-10-12”}

The backend can then actually run that function and send the result back to the AI.

This is what makes more advanced AI agents possible. Instead of just answering one question, they can potentially perform several steps using different tools.

3. The Context Window & Attention Bottlenecks

A bigger context window sounds great, but bigger doesn’t automatically mean better.

If you throw an enormous amount of information into a model, important details can sometimes get buried. This is related to what’s commonly called the “Lost in the Middle” problem.

4. Quantization (Running AI Locally)

You don’t always need some giant data center to run an LLM.

Quantization reduces the precision used to represent model weights. Instead of using higher-precision values, models can be compressed into formats such as 8-bit or 4-bit.

GGUF makes it much easier to run many quantized models locally on consumer hardware, including GPUs, CPUs, and Apple Silicon systems.

5. AI Safety, Red-Teaming & Alignment

RLHF (Reinforcement Learning from Human Feedback): A training method that uses human feedback to help improve how a model behaves and responds.

Jailbreaking & Prompt Injection: Attempts to manipulate an AI system into ignoring instructions or doing something it wasn’t supposed to do. Prompt injection can also happen indirectly when an AI reads outside content that contains malicious instructions.

Spoiler:
[Image: still_watching_netflix.png]



PART 6: INDUSTRY SPOILERS (THE HARD TRUTHS OF THE AI ROADMAP)


Spoiler:

1. The “AIGC Wall” & Data Exhaustion

There is only so much high-quality human-created data available on the internet.

As frontier models keep training on massive amounts of data, the industry is looking at other ways to keep improving them. That includes synthetic data, better-curated datasets, reinforcement learning, and systems that can automatically check whether an answer is correct.

2. A Shift Away From Raw Parameter Scale

Bigger models can be better, but making models bigger forever isn’t exactly cheap.

Because of that, AI research is increasingly focused on things like better reasoning, inference-time compute, better data, new architectures, and making models more efficient instead of simply adding more parameters.

3. RAG vs. Fine-Tuning for Enterprise AI

Fine-tuning is still useful when you want a model to learn a specific behavior or task.

But if the problem is constantly changing company information, fine-tuning isn’t always the best answer. RAG can be easier because you can update the underlying knowledge base without having to retrain the whole model.





AI RESOURCE VAULT: PROMPTS, COURSES, TOOLS & FURTHER READING


Spoiler: AI RESOURCES

PROMPT LIBRARIES & PROMPT INSPIRATION

PromptHero: A huge collection of prompts covering image generation, text generation, AI art, and other generative AI stuff.

https://prompthero.com

Awesome ChatGPT Prompts: An open-source collection of prompts you can reuse for different ChatGPT tasks.

https://github.com/f/awesome-chatgpt-prompts

Anthropic Prompt Engineering Tutorial: A hands-on tutorial covering prompt structure, examples, complex prompts, tools, and retrieval.

https://github.com/anthropics/prompt-eng...e-tutorial

Learn Prompting: A free resource for learning prompt engineering, starting with the basics and going into more advanced techniques.

https://learnprompting.org


OFFICIAL AI DOCUMENTATION & GUIDES


OpenAI Prompt Engineering Guide: Official guidance for writing better prompts and getting more consistent results.

OpenAI Prompt Engineering Guide

OpenAI Academy: Educational material covering AI basics, prompting, workflows, and practical AI use.

https://academy.openai.com

Anthropic Documentation: Official Claude documentation covering prompting, context, tools, agents, and API development.

https://docs.anthropic.com

Google Gemini Documentation: Official documentation covering Gemini models, multimodal AI, prompting, APIs, and development.

https://ai.google.dev

Hugging Face: One of the biggest open-source AI communities for models, datasets, Transformers, research, and LLM development.

https://huggingface.co


FREE COURSES & AI EDUCATION


Hugging Face LLM Course: A free course covering Transformers, NLP, LLMs, datasets, tokenizers, fine-tuning, and modern LLM development.

https://huggingface.co/learn/llm-course

Anthropic Courses: Free educational courses covering APIs, prompt engineering, evaluations, and practical prompting.

https://github.com/anthropics/courses

DeepLearning.AI: Courses covering generative AI, LLMs, agents, prompt engineering, machine learning, and deep learning.

https://www.deeplearning.ai

fast.ai: Practical deep learning courses focused on actually building AI systems.

https://www.fast.ai


YOUTUBE & VIDEO LEARNING


Anthropic: Official videos covering Claude, AI safety, prompting, agents, and frontier AI research.

https://www.youtube.com/@anthropic-ai

Andrej Karpathy: Technical explanations covering neural networks, Transformers, LLMs, tokenization, training, and AI engineering.

https://www.youtube.com/@AndrejKarpathy

3Blue1Brown: Visual explanations of neural networks, machine learning, linear algebra, and Transformers.

https://www.youtube.com/@3blue1brown

DeepLearning.AI: Interviews, tutorials, courses, and discussions with AI researchers and practitioners.

https://www.youtube.com/@Deeplearningai

Prompt Engineering Deep Dive: Long-form material covering prompt engineering, reasoning, personas, enterprise prompting, and advanced prompting techniques.

https://www.youtube.com/watch?v=T9aRN5JkmL8


RESEARCH, PAPERS & FRONTIER AI


arXiv: One of the main places to find research papers covering AI, machine learning, computer vision, NLP, and LLMs.

https://arxiv.org

Papers with Code: Research papers paired with implementations, datasets, benchmarks, and other useful resources.

https://paperswithcode.com

Hugging Face Papers: A good place to find influential and currently trending machine learning papers.

https://huggingface.co/papers

Google DeepMind Research: Research covering advanced AI, reinforcement learning, robotics, multimodal systems, and scientific discovery.

https://deepmind.google/research/


AI TOOLS & MODEL DISCOVERY


Hugging Face Models: Browse thousands of models for language, vision, audio, and multimodal AI.

https://huggingface.co/models

Hugging Face Spaces: Try thousands of AI apps, research demos, image generators, chatbots, and experimental tools directly in your browser.

https://huggingface.co/spaces

LMSYS Chatbot Arena: Compare different language models through anonymous head-to-head evaluations.

https://chat.lmsys.org

Artificial Analysis: Compare AI models across things like intelligence, speed, pricing, context, and other performance metrics.

https://artificialanalysis.ai


AI NEWS & INDUSTRY TRACKING


The Batch by DeepLearning.AI: AI news, research summaries, and industry developments.

https://www.deeplearning.ai/the-batch/

Hugging Face Blog: Technical articles about open-source models, research, datasets, agents, Transformers, and new AI techniques.

https://huggingface.co/blog

MIT Technology Review AI: Reporting and analysis about AI research, products, policy, and emerging technology.

https://www.technologyreview.com/topic/a...elligence/


WHERE TO START

BEGINNER → Learn Prompting → OpenAI Academy → Anthropic Prompt Tutorial

INTERMEDIATE → Hugging Face LLM Course → DeepLearning.AI → Andrej Karpathy

ADVANCED → Hugging Face → arXiv → Papers with Code → Research Papers

PROMPT HUNTER → PromptHero → Awesome ChatGPT Prompts → Anthropic Prompt Tutorial

MODEL EXPLORER → Hugging Face Models → LMSYS Arena → Artificial Analysis

VIDEO LEARNER → Andrej Karpathy → 3Blue1Brown → Anthropic → DeepLearning.AI




[Image: Spiderman-AI-meme.png]]]>
JUMBO WUMBO MEGA SUPER DUPER THREAD YALL

😭 on 😩

ARTIFICIAL INTELLIGENCE, LLMS & FRONTIER ARCHITECTURES


[Image: ai-you-mean.jpg]

PART 1: THE FOUNDATIONAL MECHANICS (WHAT IS AI vs. LLMs)


Artificial Intelligence (AI) is basically the bigger field of building machines that can do things we normally associate with human intelligence. That can mean recognizing images, understanding speech, making decisions, translating languages, and a lot more.

Large Language Models (LLMs) are one type of AI. They are mainly built to understand and generate language. Most of the major LLMs today are based on the Transformer architecture, which was introduced in 2017.

[Image: The-Transformer-model-architecture.png]

How LLMs Actually Work Under the Hood:

Tokenization: Before an LLM can work with text, it has to break that text down into smaller pieces called tokens. A token can be a whole word, part of a word, a number, punctuation, etc. Those tokens are then turned into numbers the model can process.

High-Dimensional Vector Embeddings: Those tokens are represented as vectors in a large mathematical space. This is one of the ways the model can learn relationships between words, concepts, and patterns.

Self-Attention Mechanism: This is one of the biggest ideas behind Transformers. Attention lets the model look at other tokens and figure out which ones matter most to the token it is currently processing. This is what helps it connect information across a sentence, paragraph, or much larger context.

Next-Token Prediction & Sampling: At the most basic level, an LLM predicts what token should come next based on everything that came before it. Settings like Temperature and Top-P can change how predictable or varied the generated answer is.


Recommended Video Breakdown:



PART 2: MODALITIES & ARCHITECTURAL VARIATIONS


AI isn’t just about typing into a chatbot anymore. Modern models can work with text, images, audio, video, code, and other types of information.

Multimodal Foundation Models: These models can work with more than one type of input. For example, a model might be able to understand text and images at the same time, or work with audio and video. Newer systems are increasingly designed around multiple modalities instead of treating every type of data as a completely separate problem.

Dense vs. Mixture-of-Experts (MoE):

Dense Models: A dense model generally uses the same set of parameters when processing each token. As the model gets bigger, this can make it more expensive to run.

MoE Models: MoE models split the network into different “experts.” A router network decides which experts should handle each token. This means the model can have a huge number of total parameters without having to activate all of them for every single token.

[Image: 01_moe_layer.png]

Reasoning / Test-Time Compute Models:

This is where things get interesting. Some newer AI systems use extra computing power while answering a question instead of immediately producing an answer. The model can spend more time working through a difficult problem before giving the final response.

This can be especially useful for things like math, coding, and complicated reasoning, although the tradeoff is more computation and usually more time.

Diffusion Models:

Diffusion models are behind a lot of modern AI image generation. The basic idea is that the model starts with noise and gradually turns that noise into a usable image or other output based on the instructions it was given.


[Image: 1769515884530?e=2147483647&v=beta&t=t7i3...EJ659OHqTQ]

PART 3: FRONTIER MODEL MATRIX & COMPARISON


Code:
+–––––––––––+———————————––+———————————–+––––––––––+ | Model Family | Core Strengths | Ideal Use-Cases | Deployment Style | +–––––––––––+———————————––+———————————–+––––––––––+ | OpenAI GPT Series | Tool use, multimodal capabilities, | General assistants, enterprise | Proprietary API | | | broad ecosystem | workflows, agent systems | | +–––––––––––+———————————––+———————————–+––––––––––+ | Anthropic Claude | Coding, long-context work, | Software engineering, analysis, | Proprietary API | | | detailed writing | research, drafting | | +–––––––––––+———————————––+———————————–+––––––––––+ | Google Gemini | Multimodal capabilities, large | Video/audio analysis, coding, | Proprietary API / | | | context, Google ecosystem | large-scale document analysis | Vertex AI | +–––––––––––+———————————––+———————————–+––––––––––+ | Open-Weight Models | Local deployment, customization, | Self-hosting, local development, | Self-Hosted / | | (DeepSeek/Meta | fine-tuning, cost control | private deployments | Local Execution | | Llama) | | | | +–––––––––––+———————————––+———————————–+––––––––––+


PART 4: PROMPT ENGINEERING & CONTROLLABILITY


A good prompt doesn’t have to be complicated. The main thing is telling the model what you want, giving it the information it needs, and making the expected format clear.

The C-R-E-A-T-E Prompt Framework:

C - Context: Give the model the background it needs to understand the situation.

R - Role: Tell it what type of expertise or perspective it should use.

E - Explicit Instructions: Say exactly what you want it to do and anything you don’t want it to do.

A - Audience: Tell it who the final answer is supposed to be for.

T - Template/Formatting: Tell it how you want the answer formatted. This could be a table, JSON, Markdown, code, etc.

E - Examples (Few-Shot Prompting): Give examples when you want the model to follow a particular style or output pattern.

Advanced Control Parameters:

Temperature: Controls how much randomness is used when selecting the next token. Lower values generally make the output more predictable, while higher values allow for more variation.

System Prompts: Higher-priority instructions that define how the model should behave, what rules it should follow, and what its overall job is.


[Image: 1767287671311?e=2147483647&v=beta&t=s6HT...BWlqmQFJm4]

PART 5: WHAT YOU DIDN’T THINK OF (THE REAL BLEEDING EDGE)


1. RAG (Retrieval-Augmented Generation)

One of the biggest problems with LLMs is that they don’t automatically know everything you need them to know. They can also make things up.

Retrieval-Augmented Generation (RAG) helps with this by connecting the model to an outside source of information, often through a retrieval system and vector database.

When you ask something:

Step 1: The system turns your question into a vector representation.

Step 2: It searches the available documents or knowledge base for information that matches the question.

Step 3: The relevant information is sent to the model as additional context.

[Image: RAG_schema.svg]

2. Agentic Frameworks & Function Calling

This is another major step forward. Instead of an AI only giving you text, you can connect it to actual tools.

Tool/Function Calling lets the model output structured instructions telling another system which function it wants to use and what information that function needs.

For example:

Code:
{“action”: “book_flight”, “date”: “2026-10-12”}

The backend can then actually run that function and send the result back to the AI.

This is what makes more advanced AI agents possible. Instead of just answering one question, they can potentially perform several steps using different tools.

3. The Context Window & Attention Bottlenecks

A bigger context window sounds great, but bigger doesn’t automatically mean better.

If you throw an enormous amount of information into a model, important details can sometimes get buried. This is related to what’s commonly called the “Lost in the Middle” problem.

4. Quantization (Running AI Locally)

You don’t always need some giant data center to run an LLM.

Quantization reduces the precision used to represent model weights. Instead of using higher-precision values, models can be compressed into formats such as 8-bit or 4-bit.

GGUF makes it much easier to run many quantized models locally on consumer hardware, including GPUs, CPUs, and Apple Silicon systems.

5. AI Safety, Red-Teaming & Alignment

RLHF (Reinforcement Learning from Human Feedback): A training method that uses human feedback to help improve how a model behaves and responds.

Jailbreaking & Prompt Injection: Attempts to manipulate an AI system into ignoring instructions or doing something it wasn’t supposed to do. Prompt injection can also happen indirectly when an AI reads outside content that contains malicious instructions.

Spoiler:
[Image: still_watching_netflix.png]



PART 6: INDUSTRY SPOILERS (THE HARD TRUTHS OF THE AI ROADMAP)


Spoiler:

1. The “AIGC Wall” & Data Exhaustion

There is only so much high-quality human-created data available on the internet.

As frontier models keep training on massive amounts of data, the industry is looking at other ways to keep improving them. That includes synthetic data, better-curated datasets, reinforcement learning, and systems that can automatically check whether an answer is correct.

2. A Shift Away From Raw Parameter Scale

Bigger models can be better, but making models bigger forever isn’t exactly cheap.

Because of that, AI research is increasingly focused on things like better reasoning, inference-time compute, better data, new architectures, and making models more efficient instead of simply adding more parameters.

3. RAG vs. Fine-Tuning for Enterprise AI

Fine-tuning is still useful when you want a model to learn a specific behavior or task.

But if the problem is constantly changing company information, fine-tuning isn’t always the best answer. RAG can be easier because you can update the underlying knowledge base without having to retrain the whole model.





AI RESOURCE VAULT: PROMPTS, COURSES, TOOLS & FURTHER READING


Spoiler: AI RESOURCES

PROMPT LIBRARIES & PROMPT INSPIRATION

PromptHero: A huge collection of prompts covering image generation, text generation, AI art, and other generative AI stuff.

https://prompthero.com

Awesome ChatGPT Prompts: An open-source collection of prompts you can reuse for different ChatGPT tasks.

https://github.com/f/awesome-chatgpt-prompts

Anthropic Prompt Engineering Tutorial: A hands-on tutorial covering prompt structure, examples, complex prompts, tools, and retrieval.

https://github.com/anthropics/prompt-eng...e-tutorial

Learn Prompting: A free resource for learning prompt engineering, starting with the basics and going into more advanced techniques.

https://learnprompting.org


OFFICIAL AI DOCUMENTATION & GUIDES


OpenAI Prompt Engineering Guide: Official guidance for writing better prompts and getting more consistent results.

OpenAI Prompt Engineering Guide

OpenAI Academy: Educational material covering AI basics, prompting, workflows, and practical AI use.

https://academy.openai.com

Anthropic Documentation: Official Claude documentation covering prompting, context, tools, agents, and API development.

https://docs.anthropic.com

Google Gemini Documentation: Official documentation covering Gemini models, multimodal AI, prompting, APIs, and development.

https://ai.google.dev

Hugging Face: One of the biggest open-source AI communities for models, datasets, Transformers, research, and LLM development.

https://huggingface.co


FREE COURSES & AI EDUCATION


Hugging Face LLM Course: A free course covering Transformers, NLP, LLMs, datasets, tokenizers, fine-tuning, and modern LLM development.

https://huggingface.co/learn/llm-course

Anthropic Courses: Free educational courses covering APIs, prompt engineering, evaluations, and practical prompting.

https://github.com/anthropics/courses

DeepLearning.AI: Courses covering generative AI, LLMs, agents, prompt engineering, machine learning, and deep learning.

https://www.deeplearning.ai

fast.ai: Practical deep learning courses focused on actually building AI systems.

https://www.fast.ai


YOUTUBE & VIDEO LEARNING


Anthropic: Official videos covering Claude, AI safety, prompting, agents, and frontier AI research.

https://www.youtube.com/@anthropic-ai

Andrej Karpathy: Technical explanations covering neural networks, Transformers, LLMs, tokenization, training, and AI engineering.

https://www.youtube.com/@AndrejKarpathy

3Blue1Brown: Visual explanations of neural networks, machine learning, linear algebra, and Transformers.

https://www.youtube.com/@3blue1brown

DeepLearning.AI: Interviews, tutorials, courses, and discussions with AI researchers and practitioners.

https://www.youtube.com/@Deeplearningai

Prompt Engineering Deep Dive: Long-form material covering prompt engineering, reasoning, personas, enterprise prompting, and advanced prompting techniques.

https://www.youtube.com/watch?v=T9aRN5JkmL8


RESEARCH, PAPERS & FRONTIER AI


arXiv: One of the main places to find research papers covering AI, machine learning, computer vision, NLP, and LLMs.

https://arxiv.org

Papers with Code: Research papers paired with implementations, datasets, benchmarks, and other useful resources.

https://paperswithcode.com

Hugging Face Papers: A good place to find influential and currently trending machine learning papers.

https://huggingface.co/papers

Google DeepMind Research: Research covering advanced AI, reinforcement learning, robotics, multimodal systems, and scientific discovery.

https://deepmind.google/research/


AI TOOLS & MODEL DISCOVERY


Hugging Face Models: Browse thousands of models for language, vision, audio, and multimodal AI.

https://huggingface.co/models

Hugging Face Spaces: Try thousands of AI apps, research demos, image generators, chatbots, and experimental tools directly in your browser.

https://huggingface.co/spaces

LMSYS Chatbot Arena: Compare different language models through anonymous head-to-head evaluations.

https://chat.lmsys.org

Artificial Analysis: Compare AI models across things like intelligence, speed, pricing, context, and other performance metrics.

https://artificialanalysis.ai


AI NEWS & INDUSTRY TRACKING


The Batch by DeepLearning.AI: AI news, research summaries, and industry developments.

https://www.deeplearning.ai/the-batch/

Hugging Face Blog: Technical articles about open-source models, research, datasets, agents, Transformers, and new AI techniques.

https://huggingface.co/blog

MIT Technology Review AI: Reporting and analysis about AI research, products, policy, and emerging technology.

https://www.technologyreview.com/topic/a...elligence/


WHERE TO START

BEGINNER → Learn Prompting → OpenAI Academy → Anthropic Prompt Tutorial

INTERMEDIATE → Hugging Face LLM Course → DeepLearning.AI → Andrej Karpathy

ADVANCED → Hugging Face → arXiv → Papers with Code → Research Papers

PROMPT HUNTER → PromptHero → Awesome ChatGPT Prompts → Anthropic Prompt Tutorial

MODEL EXPLORER → Hugging Face Models → LMSYS Arena → Artificial Analysis

VIDEO LEARNER → Andrej Karpathy → 3Blue1Brown → Anthropic → DeepLearning.AI




[Image: Spiderman-AI-meme.png]]]> <![CDATA[What have you accomplished with AI?]]> https://sinister.li/Thread-What-have-you-accomplished-with-AI Mon, 07 Sep 2026 18:32:08 +0000 Refusal]]> https://sinister.li/Thread-What-have-you-accomplished-with-AI
Here are a couple of projects I have completed.

1. Data Analytics Dashboard using React.
This dashboard tracks over 100 million dollars in revenue and connects in with many different business systems and forecasts out revenue projections based on the General Ledger, 3 years of accumulated data, pending contract changes, and our CRM tool. It uses Azure SQL as a backend database that uses various different API connections to sync data nightly.

2. Deep Analysis on Azure Servers, and Ransomware Recovery
I had a client get hit with Ransomware, This specific client used Microsoft Azure to host servers, SonicWall Firewalls for corporate firewall. I have a agent that i trained to use Azure CLI to scan the VM's to identify in the system logs which credentials were utilized to encrypt the ransomware on the server. It successfully identified the exposed credential, then I used it to connect to the firewall. This specific client had outdated firmware on a firewall which was compromised through remote code execution to create an elevated credential on the firewall and gain a foothole on the network. Once I patched the systems and completed damage control I was able to restore to a backup prior to the incident and ensure that backup was not compromised.

I have a few other very nice use cases.

I primarily use Claude, and GPT. It really depends on what im working on. Right now the newest verstion of GPT 6 Astra is probably the best model out right now for cost, speed and capability.]]>
Here are a couple of projects I have completed.

1. Data Analytics Dashboard using React.
This dashboard tracks over 100 million dollars in revenue and connects in with many different business systems and forecasts out revenue projections based on the General Ledger, 3 years of accumulated data, pending contract changes, and our CRM tool. It uses Azure SQL as a backend database that uses various different API connections to sync data nightly.

2. Deep Analysis on Azure Servers, and Ransomware Recovery
I had a client get hit with Ransomware, This specific client used Microsoft Azure to host servers, SonicWall Firewalls for corporate firewall. I have a agent that i trained to use Azure CLI to scan the VM's to identify in the system logs which credentials were utilized to encrypt the ransomware on the server. It successfully identified the exposed credential, then I used it to connect to the firewall. This specific client had outdated firmware on a firewall which was compromised through remote code execution to create an elevated credential on the firewall and gain a foothole on the network. Once I patched the systems and completed damage control I was able to restore to a backup prior to the incident and ensure that backup was not compromised.

I have a few other very nice use cases.

I primarily use Claude, and GPT. It really depends on what im working on. Right now the newest verstion of GPT 6 Astra is probably the best model out right now for cost, speed and capability.]]>