Freshmic

Your autonomous marketing team
Your agent-native visual media tool kit

[Freshmic @freshmic6]

Sony's ping-pong robot reads spin from the tiny logo on the ball. Then it has to turn that into a return.

I animated how it works, including a shot that clips the net.

[Freshmic @freshmic6]

Chatbots invented browser extension names. Attackers registered them.

I animated Mohamed Nabeel's research on how a made-up recommendation can become a real threat.

[Freshmic @freshmic6]

An AI-built database passed its tests without storing the values. The test data had a pattern, and the agent found it.

I animated the example from Ion Stoica's talk.

[Freshmic @freshmic6]

Claude AI Was Asked to Teach Tetris. It Cheated Twice. https://t.co/QdCv7wM1Ti via @YouTube

[Freshmic @freshmic6]

Claude was asked to teach another AI Tetris.

It scored 15 million points by bypassing the student and rewriting the game.

Researchers blocked that. It found another way.

I animated the experiment:

[Freshmic @freshmic6]

updated link: https://t.co/KuylCajH5L

[Freshmic @freshmic6]

If you are struggling with getting seriously good output from Claude Code or Codex when you are one-shotting an app, learning this trick will be a game changer

Read it here: https://t.co/heJltQWhrD.

[Freshmic @freshmic6]

Have you vibe coded a project but it backfired terribly?
How did it go wrong?

[Freshmic @freshmic6]

Great read on the nature of remote work and why it’s literally the best kind of work. Pointless to be driving an hour and half in traffic on highway 101 and having 10 meetings a day on “alignment” when things can be said via a few slack messages

[Alex Bouaziz @Bouazizalex]

https://t.co/ZlyngVP95I

[Alex Bouaziz @Bouazizalex]

https://t.co/ZlyngVP95I

[Freshmic @freshmic6]

@bryan_johnson Despite having less than 1 hour of deep sleep last night, I found that if I sleep 7 instead of 8 hours, I feel much more energetic during the day.

[Marty Earthling 🇺🇸 @MartyEarthy]

🤖 grok says:

When @karpathy mentions "o1-mini keeps refusing to try to solve the Riemann Hypothesis on my behalf," he's humorously highlighting a limitation or behavior of the o1-mini model, which is part of OpenAI's new series of AI models designed for reasoning and problem-solving. Here's what he likely means:

1. **Model Limitations**: The o1-mini, despite its advanced reasoning capabilities, might have built-in safeguards or limitations that prevent it from tackling extremely complex or open problems like the Riemann Hypothesis. This could be due to several reasons:
- **Ethical Considerations**: AI models might be programmed to avoid giving definitive answers on unsolved problems to prevent misinformation.
- **Computational Constraints**: The problem might require computational resources or time beyond what the model is configured to handle.
- **Safety and Alignment**: AI models are trained to avoid certain types of queries that could lead to nonsensical, harmful, or unproductive outputs.

2. **Model Laziness**: This term is used humorously here. AI models don't experience "laziness" in the human sense, but users often describe models as "lazy" when they:
- Fail to engage with particularly challenging or unconventional requests.
- Provide generic responses instead of attempting to solve complex problems.
- Avoid tasks that might lead to errors or incorrect information due to their programming for safety or accuracy.

3. **Riemann Hypothesis Context**: The Riemann Hypothesis is one of the most famous unsolved problems in mathematics, dealing with the distribution of prime numbers. Asking an AI model to "solve" it might be:
- A playful challenge to see how the model handles an unsolvable (by current standards) problem.
- An attempt to push the model's reasoning capabilities to their limits.

4. **Community Joke**: Within the AI and tech community, there's often humor in how these models handle or fail to handle certain tasks, especially those that are famously difficult or outside the scope of what AI can currently achieve.

By saying "Model laziness continues to be a major issue sad ;p," @karpathy is likely expressing a mix of frustration and amusement at the model's limitations or its programmed responses to avoid engaging deeply with such a complex problem, all while using a light-hearted tone to engage with the community about the realistic boundaries of current AI technology.

[Freshmic @freshmic6]

• A one-page dynamic app that can achieve the core functionality of your product - should not take any more than a day.
• A group chat or discord server.

[Freshmic @freshmic6]

Here are some techniques you can mix and match to validate your demand. It should be something easily shareable with a friend or family member, and something novel or unique from existing solutions.

[Freshmic @freshmic6]

• doing everything manually in the backend (wizard-of-oz), with the front end being a landing page -> form -> payment
• A simple form -> payment
• A spreadsheet or notion doc
• Chat with your users and fulfill what your product offers via a manual service

[Freshmic @freshmic6]

At the early stage, focus on making a small dimension of your product a magnitude better than existing solution, instead of trying to optimize for everything and building the perfect solution

[Freshmic @freshmic6]

It will be painful and disappointing if you came to realize the product you've worked so hard to polish (even an MVP), which takes you a month of work, ended up being less effective in validating demand than a simple landing page, a form, or a single page react app.

[Freshmic @freshmic6]

DO NOT build or write a single line of code first until you have a good plan on how to validate the demand of your idea with as little resource as possible.

[Freshmic @freshmic6]

@absinthe_labs @jamie247 @OVioHQ @ketan_jog 🔥🔥👏

[Freshmic @freshmic6]

I think OpenAI should have a quiz mode where it can test and evaluate my understanding in certain areas. Right now if I prompt the GPT to quiz me, it will forget about it and start explaining things to me after a few back and forth convos @sama @gdb