Kicka

Captain’s Log.

[Kicka @k111cka]

on the board https://t.co/uKDcQRQQXZ #megadev

[Kicka @k111cka]

@amoamoamo @greg3z dear @amoamoamo any chance for one more spot?

[Matt Pocock @mattpocockuk]

My skills repo has 160K stars, 7.5m downloads...

...and no tutorial.

So, here it is. Watch me walk through the essential skills:

- /grill-with-docs
- /to-spec
- /to-tickets
- /implement
- /code-review

It's the whole flow, end-to-end. Enjoy:

[Claude @claudeai]

Fable 5 is back.

[Ethan Mollick @emollick]

Claude Code: "make a really amazing horror game based on the poetry of William Carlos Williams, just the wheelbarrow and the plums poems."

All the writing and design by Claude. I actually found it unnerving, despite the crude, "hand-drawn" graphics. https://t.co/yFrXgfLNMl

[swyx @swyx]

some amount of ego here but v interesting metaintelligence display where a 22 yo is selfaware enough that she can direct her own growth in the direction she wants and is unapologetically proud / excited to do so

continual learning be damned, this is self directed learning that most humans arent even aware they are capable of.

i dont know about the neuroplasticity argument - i found myself capable of a big pivot at age 30 and bet i could still do it today to a lesser degree. you can just will yourself into a different mindset.

[Paul Graham @paulg]

Prediction: In the AI age, taste will become even more important. When anyone can make anything, the big differentiator is what you choose to make.

https://t.co/3GQUlfH58t

[Simon Willison @simonw]

@bigmrrobert @steipete If you're an autodidact - someone who thrives on learning things independently - LLMs are a gift from the heavens

The rate at which I've been learning new things and new skills has accelerated SO much in the last three years thanks to assistance from LLMs

[Kicka @k111cka]

@rom1trs autopilot

[Adam @_overment]

my TOP10 @cursor_ai and @claudeai code techniques:

#1 ask me anything

instead of asking the agent to do something, point at files you know are relevant and ask a question such as 'explain to me how this works, so I can see you understand the code'. and when you feel it understands, move to the actual task.

#2 time travel

split the conversation in the middle, in a place where the context quality was good. this way you can try another path or switch to the next task while having information that is relevant to the task

#3 body swap

ask the model to switch the roles by saying something like this "okay, if you were me, what would you ask for to get [desired results]?" this way you often get high-quality instruction describing the issue and asking for things you wouldn't ask yourself. copy the response, and do "time travel" by replacing your request with the generated prompt.

#4 look at me!

agents are eager to make edits right away, so add something like "Please do not edit any files yet. Answer me directly so I can see that you understand the task and that we have everything needed to proceed." to the beginning AND the end of your prompt -- put this sentence into your text expander. then discuss the issue with the agent and when you feel it's ready, let it make the changes.

#5 mastermind group

run multiple chats on the exact same issue and ask them to write down their observations in a separate .md file. when they finish, pick the best model available and ask for synthesis. then based on it, implement a solution.

#6 rephrase

add this to the end of your message: "before moving on, please rephrase what i just said so i can clearly see you understand what i mean. include file references, function names, variables, patterns, techniques and anything you consider relevant to understand what i say." and again copy the agent's response and replace your message with whatever it says (make edits if needed)

#7 bullet time

use screen recording software to get video of the interaction and extract frames to show UI-related issues that are hard to describe. ensure you work with a model that has good taste and frontend skills such as Sonnet 4.5.

#8 find patterns

current LLMs are more knowledgeable than any of us, and we can use them as leverage. when solving a problem, ask for naming patterns, techniques, algorithms, theorems, tools, libraries, and other resources that could be useful for your task. have the AI generate a list of names with descriptions or examples so you can identify which ones seem relevant enough to explore further. be cautious of hallucinations here.

#9 logger

when facing an issue that an LLM cannot solve in 2–3 turns, ask it to add logs to track relevant variables. in web programming, `console.time` is especially useful for debugging performance issues. for readability, use tables. to track different types of logs, use prefixes so you can easily filter what you need.

#10 that's the issue

for complex issues, after analyzing the code, ask the model to write a GitHub-issue-like document that describes the task for another programmer who won't see the ongoing conversation so we need to include all the necessary context. iterate over this document a couple of times asking various questions that double-check if the document is complete and noise-free

--

what are yours?

[Adam @_overment]

I build product that integrates @OpenAI @AnthropicAI @xai @GoogleDeepMind @deepseek_ai @perplexity_ai and @GroqInc.

Here's what I've learned so far:

note: These are mostly challenges, with some opportunities.

1. The API changes very quickly and becomes inconsistent across models from the same provider. Property changes, such as switching from 'max_tokens' to 'max_completions_tokens' make mapping extremely difficult.

2. The API limitations of the models are significant. Function calling, the web search tool, and structured outputs are supported only by specific models, which complicates the logic. More importantly, this confuses users who expect all features to work with any model and end up blaming the app rather than the provider.

3. Shared responsibility rests with both providers and developers, and both contribute to building user trust. An unstable API, quantization, routing, and other "optimization techniques" undermine trust in the app, while poor implementation undermines trust in the models.

4. Competition that has made research and most specifications closed makes building on top of LLMs much harder, since the model is not just a “black box” but also a “black box hidden in a closed room.” An example of this is tokenization, about which we know very little, and counting tokens for multimodal data is extremely difficult and often requires dirty hacks.

5. All state-of-the-art models are now roughly the same; you can think of them as people with different personalities. At present, most of the logic works (and fails) in similar ways across models, but the best results come from combining them and having them work on the same issue.

6. Reasoner models are slow, costly, and difficult to steer. As a developer, I find them useful for end users but not as practical for building products on top of them, and by this I mean reasoner models in their current state. At the same time, I appreciate that we still see non-reasoner models.

7. Services like OpenRouter or Groq are helpful, but their APIs never fully match the exact implementation of the original provider. This forces developers to support both their APIs and the original provider's API, which complicates the logic and cost estimations. At the same time, OpenRouter, Groq, and similar tools are needed and important.

8. Open source models do not deliver on their promise for end users. Their marketing raises expectations to the extreme, and users expect these models to be readily available. When the experience turns out to be poor, they place the blame on the developers. With the latest models, there is a visible trend of change, but the gap between commercial models still remains.

9. Building anything that goes beyond a chat UI without funding is currently nearly impossible because of two factors: price and speed. Complex logic increases both latency and costs. Achieving results that justify the waiting time is really, really hard.

10. Censorship (I’m not sure if this is the right word) and alignment are very important, but both face challenges in defining a clear boundary between safety and usefulness. I’d say that most of the time, the model or API refuses to do something harmless while still allowing potentially harmful things.

conclusions:

- big labs should take care to provide an API that is consistent within their own frameworks and well-designed.
- providers such as OpenRouter should offer as much compatibility with all providers as possible, and it would be nice if big labs helped them achieve this.
- libraries such as AI SDK are probably the most important tools we can build, as they allow developers to focus on creating solutions on top of LLMs rather than struggling with various APIs.

It's still very early.

[Ethan Mollick @emollick]

I think OpenAI missed a bit of an opportunity to show GPT-4.5’s strengths, to their detriment & to the AI industry as a whole by only using the same coding & test benchmarks when critical thinking & ideation are key AI use cases where 4.5 is good.

Those are actually measurable

[Alex Albert @alexalbert__]

Introducing Claude 3.7 Sonnet.

Our most intelligent model to date and the first generally available hybrid reasoning model in the world.

[Kicka @k111cka]

@levelsio How is Starlink working for you? I need good internet for one month while I am abroad. Is it better to rent Starlink equipment than to buy a SIM card for internet?

[Kicka @k111cka]

@zacbray_ InfoPlus

[Kicka @k111cka]

@RMHildebrandt send

[Andrej Karpathy @karpathy]

⚡️ Excited to share that I am starting an AI+Education company called Eureka Labs.
The announcement:

---
We are Eureka Labs and we are building a new kind of school that is AI native.

How can we approach an ideal experience for learning something new? For example, in the case of physics one could imagine working through very high quality course materials together with Feynman, who is there to guide you every step of the way. Unfortunately, subject matter experts who are deeply passionate, great at teaching, infinitely patient and fluent in all of the world's languages are also very scarce and cannot personally tutor all 8 billion of us on demand.

However, with recent progress in generative AI, this learning experience feels tractable. The teacher still designs the course materials, but they are supported, leveraged and scaled with an AI Teaching Assistant who is optimized to help guide the students through them. This Teacher + AI symbiosis could run an entire curriculum of courses on a common platform. If we are successful, it will be easy for anyone to learn anything, expanding education in both reach (a large number of people learning something) and extent (any one person learning a large amount of subjects, beyond what may be possible today unassisted).

Our first product will be the world's obviously best AI course, LLM101n. This is an undergraduate-level class that guides the student through training their own AI, very similar to a smaller version of the AI Teaching Assistant itself. The course materials will be available online, but we also plan to run both digital and physical cohorts of people going through it together.

Today, we are heads down building LLM101n, but we look forward to a future where AI is a key technology for increasing human potential. What would you like to learn?
---

@EurekaLabsAI is the culmination of my passion in both AI and education over ~2 decades. My interest in education took me from YouTube tutorials on Rubik's cubes to starting CS231n at Stanford, to my more recent Zero-to-Hero AI series. While my work in AI took me from academic research at Stanford to real-world products at Tesla and AGI research at OpenAI. All of my work combining the two so far has only been part-time, as side quests to my "real job", so I am quite excited to dive in and build something great, professionally and full-time.

It's still early days but I wanted to announce the company so that I can build publicly instead of keeping a secret that isn't. Outbound links with a bit more info in the reply!

[WikiLeaks @wikileaks]

JULIAN ASSANGE IS FREE

Julian Assange is free. He left Belmarsh maximum security prison on the morning of 24 June, after having spent 1901 days there. He was granted bail by the High Court in London and was released at Stansted airport during the afternoon, where he boarded a plane and departed the UK.

This is the result of a global campaign that spanned grass-roots organisers, press freedom campaigners, legislators and leaders from across the political spectrum, all the way to the United Nations. This created the space for a long period of negotiations with the US Department of Justice, leading to a deal that has not yet been formally finalised. We will provide more information as soon as possible.

After more than five years in a 2x3 metre cell, isolated 23 hours a day, he will soon reunite with his wife Stella Assange, and their children, who have only known their father from behind bars.

WikiLeaks published groundbreaking stories of government corruption and human rights abuses, holding the powerful accountable for their actions. As editor-in-chief, Julian paid severely for these principles, and for the people's right to know.

As he returns to Australia, we thank all who stood by us, fought for us, and remained utterly committed in the fight for his freedom.

Julian's freedom is our freedom.

[More details to follow]

[Kicka @k111cka]

"All those moments will be lost in time" https://t.co/MOJRIxuTjI

[Kicka @k111cka]

Dr. Dolittle is here
https://t.co/xSb57vZGjU