CEO of NewTech Prime Co., Ltd.; Author of "Real Software Modeling" and "Learning UML2 While Thinking"; Developer of MyAISmarteasy.
Seoul, Korea
Joined April 2010
- Tweets2.8K
- Following540
- Followers576
- Likes14
김현남 (Justine Kim) retweeted
Not all AI agents are built the same. So what sets them apart?
Here’s a breakdown of 10 core types of AI agents you’ll come across in real-world systems, from simple reactive agents to complex multi-agent systems.
1. Task-Specific AI Agent
Built for one focused task like summarizing or translating. It follows a fixed process with no learning or adaptation.
2. Reactive Agent
Responds to immediate input without using memory or history. Think of it like a reflex - it reacts, not plans.
3. Model-Based Agent
Builds an internal map of its environment. Simulates outcomes before acting to make smarter, context-aware decisions.
4. Goal-Based Agent
Starts with a goal and works backward. It plans steps, simulates paths, and selects the route that achieves the goal.
5. Utility-Based Agent
Chooses actions based on how beneficial they are. It weighs all options and picks the one with the highest value.
6. Learning Agent
Improves over time by learning from past actions. Adjusts its strategy using feedback and stores new knowledge.
7. Planning Agent
Focuses on long-term strategy. It defines a goal, maps out steps, and adjusts based on progress not just reaction.
8. Reflex Agent with Memory
Uses preset rules but with added memory of past inputs. Helps respond better when situations repeat or evolve.
9. Multi-Agent System Agent
Works with or against other agents. They share environments, negotiate roles, and coordinate to reach a bigger goal.
10. Rational Agent
Always selects the most logical option. It analyzes the full picture, predicts outcomes, and chooses the smartest path.
Save this if you're exploring Agentic AI or designing intelligent decision-making systems.
김현남 (Justine Kim) retweeted
I used to think typing faster meant getting more done.
But it only made me feel more tired and distracted.
Every week, I spent hours writing emails, fixing drafts, and trying to make every line sound perfect.
It wasn’t productive work, it was just overthinking on repeat.
Then I started using Typeless / @typelessdotcom and everything felt lighter.
Now I simply speak my thoughts and Typeless turns them into clean, ready-to-use text within seconds.
✅ It makes client emails sound sharp.
✅ It creates blog posts, proposals, or even AI video scripts that look perfectly formatted.
✅ It can even translate content into any language just by saying translate to Spanish and the result appears beautifully formatted.
What I love most is how it feels like a creative partner that keeps up with my mind.
It helps me stay focused, think clearly, and express ideas without friction.
Typing used to slow me down.
Now my thoughts move straight into words with Typeless.
Try it at typeless.com
#Typeless #WritingTools #WorkSmarter #VoiceToText
김현남 (Justine Kim) retweeted
GPT-5 → Best for logic & reasoning
Claude 4.5 → Best for writing & memory
Gemini 2.5 → Best for research + images
Perplexity → Best for instant answers
Runway → Best for cinematic AI video
I just found a secret trick to use all of them at one place.
Here’s how ↓
김현남 (Justine Kim) retweeted
Just focus on something for 4 hours every day, you end up building something insanely impressive.
Below is a simple AI roadmap you can follow.
One small task per day.
Ship something every week.
Start using these tools for 30 days:
> For tasks
• ChatGPT
• Gemini
• Claude
• Grok
> For Vibe coding
• Cursor
> For AI video
• Kling
• Higgsfield
• Vibepost. app
4 hours/day. Ship weekly. Repeat.
김현남 (Justine Kim) retweeted
99% of people still use plain text to prompt AI.
That’s why their results are random, messy, or wrong.
Use JSON, Markdown, or XML - and the model does exactly what you want.
Here’s a quick guide (with examples): 🧵
김현남 (Justine Kim) retweeted
🎤 CARV KOREA AMA
📅 7월 10일 (목) | ⏰ 한국시간 밤 10시
💰 총 상금: 100 $CARV + 최대 500 GEM
🎮 참여 링크:
👾 플레이어 (300 GEM)
🔗 play.carv.io/events/9c0c4f04…
👥 청중 (200 GEM)
🔗play.carv.io/events/fdcd2ca8…
함께 해요!
@carv_official
김현남 (Justine Kim) retweeted
We're missing (at least one) major paradigm for LLM learning. Not sure what to call it, possibly it has a name - system prompt learning?
Pretraining is for knowledge.
Finetuning (SL/RL) is for habitual behavior.
Both of these involve a change in parameters but a lot of human learning feels more like a change in system prompt. You encounter a problem, figure something out, then "remember" something in fairly explicit terms for the next time. E.g. "It seems when I encounter this and that kind of a problem, I should try this and that kind of an approach/solution". It feels more like taking notes for yourself, i.e. something like the "Memory" feature but not to store per-user random facts, but general/global problem solving knowledge and strategies. LLMs are quite literally like the guy in Memento, except we haven't given them their scratchpad yet. Note that this paradigm is also significantly more powerful and data efficient because a knowledge-guided "review" stage is a significantly higher dimensional feedback channel than a reward scaler.
I was prompted to jot down this shower of thoughts after reading through Claude's system prompt, which currently seems to be around 17,000 words, specifying not just basic behavior style/preferences (e.g. refuse various requests related to song lyrics) but also a large amount of general problem solving strategies, e.g.:
"If Claude is asked to count words, letters, and characters, it thinks step by step before answering the person. It explicitly counts the words, letters, or characters by assigning a number to each. It only answers the person once it has performed this explicit counting step."
This is to help Claude solve 'r' in strawberry etc. Imo this is not the kind of problem solving knowledge that should be baked into weights via Reinforcement Learning, or least not immediately/exclusively. And it certainly shouldn't come from human engineers writing system prompts by hand. It should come from System Prompt learning, which resembles RL in the setup, with the exception of the learning algorithm (edits vs gradient descent). A large section of the LLM system prompt could be written via system prompt learning, it would look a bit like the LLM writing a book for itself on how to solve problems. If this works it would be a new/powerful learning paradigm. With a lot of details left to figure out (how do the edits work? can/should you learn the edit system? how do you gradually move knowledge from the explicit system text to habitual weights, as humans seem to do? etc.).
김현남 (Justine Kim) retweeted
STUDY SMARTER WITH CHATGPT
Here are 7 prompts to learn faster and retain more:
김현남 (Justine Kim) retweeted
Agent2Agent Protocol vs. Model Context Protocol, clearly explained (with visual):
- Agent2Agent protocol lets AI agents connect to other Agents.
- Model context protocol lets AI Agents connect to Tools/APIs.
Both are open-source and don't compete with each other!
김현남 (Justine Kim) retweeted
Grok 3 is a GENIUS stock trader.
But most people don't know how to use it.
Here are 10 prompts to use for stock trading automation:
김현남 (Justine Kim) retweeted
You can now do deep research for free.
Google’s Gemini just quietly leveled up its free tier.
It’s crawling thousands of websites and generating 20+ page reports.
12 INSANELY valuable prompts for making money 🧵:
김현남 (Justine Kim) retweeted
o3 BEATS R1 OVERALL AND BLOWS EVERYONE ELSE AWAY IN CODING
o3-mini high became the BEST LLM BY FAR when it comes to a combination of performance, speed, and price
- beats o1, Sonnet, and others BY A LOT in coding
- 2x cheaper than Sonnet and 15x cheaper than o1
- ~5x faster than R1
- 2nd best model right after o1 in all categories
ChatLLM and CodeLLM now have o3-high if you want to play with it.
김현남 (Justine Kim) retweeted
Richard Sutton, father of Reinforcement Learning, just gave a great talk where he said "decentralized collaboration is the source of all human flourishing and the source of all that's good in the world." And if you're in favor of this, "the enemy" are the people calling for "centralized control" of both "people and AI."
Centralized control of AI is the same as calls for centralized control of people. Control of AI chips and who can download what and what AI's goals are allowed to be, is the same as calls to limit free speech, and dictating who can trade with who, and who you can collaborate with, and what you are allowed to think.
Authoritarians are the parasites of any system. They are the selfish cheaters who try to stack the benefits for themselves and their little tribe and they cause nothing but pain and misery for themselves and others.
Trade, art, economics, community come from individual agents/people pursuing their own individual goals and choosing to work together. That is the light in the world.
The darkness is the control freaks and the parasitic cheaters who try to stack the deck in their favor. Their worldview is one of darkness, of us versus them, and their fundamental belief is most people are evil and can't be trusted.
Make no mistake, calls to restrict AI, to align AI, to create kill switches in chips, and to limit access to chips, and to embedded surveillance trackers in chips, are the same as calls to limit speech, and control where you go, and who you can work with and do business with. And these calls only ever come from the parasites of a system, the people who poison a healthy working system and who periodically pop up to set back the great and powerful and inevitable surge of human progress pressing forward endlessly and forever.
(Link to talk in comments)
김현남 (Justine Kim) retweeted
10 Grok-3 Prompts that will turn you into a Superhuman:
(Save it for later)
김현남 (Justine Kim) retweeted
Master Web3: Blockchain, Cryptocurrencies, NFTs, and the Metaverse for Free.
Join the world's 1st Web3 MOOC taught by industry experts.
Earn a blockchain-verifiable certificate.
김현남 (Justine Kim) retweeted
Claude 3.7 Sonnet with Claude Code creates an entire
"glass like" design system in one shot, with ALL the components.
How insane is this?
김현남 (Justine Kim) retweeted
The AI-Co Scientist paper dropped! The @GoogleAI -Co Scientist is a Multi-Agent System build with @GoogleDeepMind Gemini 2.0 designed to generate scientific hypotheses through generate, debate, and evolve approach combining specialized agents and reasoning. The 80 page long paper includes prompts, example outputs and full reports generated by the AI-Co Scientist. 👀
Agents:
> Supervisor Agent: Orchestrates the research process, assigns tasks to agents, allocates resources to other agents based on the research plan.
> Generation Agent: Creates initial research hypotheses by exploring literature, simulated debates, and identifies testable assumptions.
> Reflection Agent: Reviews hypotheses (peer review), assesses correctness, quality, novelty, and potential to explain existing observations, with different review types.
> Ranking Agent: Creates pairwise comparisons of hypotheses (tournament-style), using simulated debates to create a Elo rating
> Evolution Agent: Refines best hypotheses by grounding them in literature, improving coherence/feasibility, combining ideas and exploring "out-of-the-box" thinking.
> Proximity Agent: Groups similar hypotheses to optimize exploration diversity.
> Meta-review Agent: Synthesizes insights from all reviews, identifies patterns, optimizes other agents' performance, and creates report
Implementation:
1️⃣ Scientist provides a research goal, e.g., "Find AML drug candidates".
2️⃣ Supervisor Agent parses the research goal and creates a structured research plan.
3️⃣ Generation Agent explores literature and engages in simulated debates to generate initial hypotheses.
4️⃣ Reflection Agent identifies weaknesses via 6 review types (e.g., deep verification).
5️⃣ Ranking Agent creates pairwise Elo tournaments through AI debates.
6️⃣ Evolution Agent refines the top-ranked hypotheses.
7️⃣ Meta-review Agent generates overview report and shares it with the Scientist.
8️⃣ Scientist provides feedback, adds hypotheses for the next iteration.
Insights:
🧠 Powered by Google Deepmind Gemini 2.0 models
💡 Using Thinking/Reasoning Model boosted hypothesis quality by 300+ Elo points
📈 Specialized agents, mimicking the scientific method, imporved hypothesis generation
🔄 Self-improving loop: Meta-review feedback increased hypothesis novelty scores by 27%
👥 Expert-in-the-loop design allows scientists to refine goals and provide feedback
📝 Design of the prompts are a significant factor for performance
👀 Paper includes prompts, example outputs and full reports generated
🦠 Rediscovered (in 2 days) a bacterial gene transfer mechanism (took 10 years of manual) -> not sure thats a fair comparison
