likhit

AI | Bibliophile

[likhit @likhitjuttada]

@ethansteininger @grok what is v5?

[likhit @likhitjuttada]

It's an iron rule of history that every imagined order doesn't acknowledge it's fictional origins and claims to be natural and inevitable.

[likhit @likhitjuttada]

Let's go!!

[likhit @likhitjuttada]

An imagined order can be maintained only if large segments of the population and in particular large segments of the elite truly believe in it

[likhit @likhitjuttada]

Claude Code resolves /name by looking for a SKILL.md in skills directory or an <name>.md file in the older flat file commands directory and loads its instructions into the conversation, and Claude executes them as a turn.

[likhit @likhitjuttada]

Under the hood, slash commands are one of 3 things:
1. Built-in: (e.g /model, /clear): behavior encoded directly into the CLI
2. Prompt based skills: these are essentially a prompt handed to Claude
3. Workflows: Claude writes a script and a runtime executes it in the background

[likhit @likhitjuttada]

@NielsRogge @claudeai @harshagundal @huggingface Thanks for the diagram. How will we handle the case where 2 keys of the input schema have one or more overlapping values

{"risk_level": HIGH | LOW,
"income_level": HIGH | LOW"}

[likhit @likhitjuttada]

The essence of Agricultural revolution: the ability to keep more people alive under worse conditions

[likhit @likhitjuttada]

"Don't believe tree-huggers who claim that our ancestors lived in harmony with nature"

[likhit @likhitjuttada]

Reasoning effort directly correlates with answer length, a form of inference time scaling

[likhit @likhitjuttada]

At a fundamental level, almost all reasoning performance improvements boil down to spending more compute/tokens on the problem—either before deployment (during training) or runtime (during inference).

[likhit @likhitjuttada]

Notes on Dropless MoE
- expert capacity is similar to batch size but its chosen dynamically per token by the router
- having less capacity per expert leads to dropped tokens
- more than needed leads to padding & wasted compute
- Block sparse matrix multiplication to the rescue

[likhit @likhitjuttada]

"Building in our discoveries only makes it harder to see how the discovering process can be done."

https://t.co/Lcwpx31ixH

[Addy Osmani @addyosmani]

https://t.co/qRftL7VguM

[likhit @likhitjuttada]

Best explanation of loop engineering!

[ClaudeDevs @ClaudeDevs]

https://t.co/v694m8Eaj6

[likhit @likhitjuttada]

@GergelyOrosz This write-up explains it well!

[ClaudeDevs @ClaudeDevs]

https://t.co/v694m8Eaj6

[likhit @likhitjuttada]

"When you're unhappy with a result, check the context before you touch model or effort.

If Claude still gets it wrong, ask: did it not know enough, or did it not try hard enough? Not knowing enough is a model problem, not trying hard enough is an effort problem."

[ClaudeDevs @ClaudeDevs]

https://t.co/dxoI9Nna86

[Geoffrey Litt @geoffreylitt]

Hot take: I think it's still important to understand the code that our agents write!

In this mega thread (based on my AIE talk today), I will explain why that's the case, and show some ideas for how to efficiently understand code. Alright, let's dive in. 1/