Matt Sims

Founder at Better Engines. Previously: Head of Applied AI at Spotter, Machine Learning at Sudowrite, Visiting Scientist at Netflix, and Postdoc at UC Berkeley.

[Matt Sims @mbwsims]

Teach Claude Code to think systematically.

I got tired of having the same conversation with Claude Code. Review this for security. Are these tests sufficient? Can you find patterns in my codebase and update the instruction files? The answers were ok but inconsistent: no clear methodology, no memory between sessions, no systematic depth.

So I built one Claude Code plugin, then another. Before I knew it I had five, covering instruction files, test coverage, security, codebase analysis, and code evolution. I decided to merge them into one integrated plugin. Claude universe was available so I figured why not…

The Claude Universe plugin: teach Claude Code to think systematically

Github link (entirely open source): https://t.co/kGFSGnryRu

More at https://t.co/KLovBt5vDI

[Matt Sims @mbwsims]

@pbakaus Wow, this looks very awesome

[Paul Bakaus @pbakaus]

Design inside your codebase.

Introducing Impeccable 3.0:
▸ 1 skill, self-contained, 23 commands
▸ /impeccable live: pick in-browser, get prod-grade variants, accept writes to *source*
▸ reads+writes DESIGN.md + PRODUCT.md
▸ brand & product design

https://t.co/x2KG2Fjjkt

[Matt Sims @mbwsims]

It took 8 years from this post (thanks to some major advancements in LLMs) for this to become easy to implement. But the bigger takeaway is that there's a ton of theoretical concepts in computing and technology from the last century that are now almost effortless to build. We should all be doing a better job of reading those archives. They're full of some pretty incredible ideas.

[Andrej Karpathy @karpathy]

"As We May Think" Vannevar Bush in 1945 trying to predict future https://t.co/JBqyDAPYTo "A memex is a device in which an individual stores all his books, records, and communications, [...] it may be consulted with exceeding speed and flexibility. [...] supplement to his memory."

[Paul Bakaus @pbakaus]

Matt recently showed me what he’s been tinkering on behind closed doors and he started to build in the open and I was impressed!

Very excited to see his projects drop, many at the intersection of creativity, storytelling and AI. You should def give him a follow if you’re into narrative storytelling (he’s got an interesting background - English PhD *and* ML/startup background)

[Matt Sims @mbwsims]

@PeterHollens Break it and then open a PR!

[Matt Sims @mbwsims]

Teach Claude Code to think systematically.

I got tired of having the same conversation with Claude Code. Review this for security. Are these tests sufficient? Can you find patterns in my codebase and update the instruction files? The answers were ok but inconsistent: no clear methodology, no memory between sessions, no systematic depth.

So I built one Claude Code plugin, then another. Before I knew it I had five, covering instruction files, test coverage, security, codebase analysis, and code evolution. I decided to merge them into one integrated plugin. Claude universe was available so I figured why not…

The Claude Universe plugin: teach Claude Code to think systematically

Github link (entirely open source): https://t.co/kGFSGnryRu

More at https://t.co/KLovBt5vDI

[Paul Bakaus @pbakaus]

Introducing Impeccable 2.0.

• data-driven skill rewrite (evals across 7 niches) → better font/color diversity
• /critique: subagent de-bias + deterministic anti-pattern detection
• visual mode: /critique, CLI, (soon) Chrome
• npx impeccable detect (files + URLs)

Demo:

[Matt Sims @mbwsims]

Some valuable research on LLMs, fiction writing, and how to measure creativity: "One of the main contributions of our work is the collection of 14 tests, referred to as the Torrance Test for Creative Writing (TTCW), to evaluate creativity in short fictional stories."

[Tuhin Chakrabarty @TuhinChakr]

Can #GPT4 ever write fiction that matches the quality of @NewYorker fiction? Bothered by claims about AI surpassing human creativity🤔?
Good news🥁:AI is still 3-10X worse at creativity based on our rubric "Torrance Tests for Creative Writing” #NLProc #HCI
https://t.co/yp8MBpFJ9f

[Xian Li @xl_nlp]

We were wondering: can we build a good instruction LLaMa W/O relying on large amounts of human annotations or distillation from other models?

We found a scalable recipe where the model itself can be put in the loop of generating and curating finetuning data to further self-train

[Jerry Wei @JerryWeiAI]

New @GoogleAI paper! 📜

Language models repeat a user’s opinion, even when that opinion is wrong. This is more prevalent in instruction-tuned and larger models.

Finetuning with simple synthetic-data (https://t.co/CSfoZw3qOL) reduces this behavior.

https://t.co/Tux2LCs4Nl

1/

[Matt Sims @mbwsims]

Highly recommend submitting to the AI Song Contest. This was our submission from 2021: https://t.co/iJl9rFdATm

[AI Song Contest @aisongcontest]

🏆 Become the 2023 winner! How to participate?

1️⃣ Make the best AI song ever with ethics and responsability
2️⃣ Fill the form telling us your human-AI co-creation process
3️⃣ Submit your entry by September 4 👉🏼 https://t.co/Ld8qm9Rzi5

Moita sorte 🎉 #AISongContest #WalkingWithAI

[Emma Pierson @2plus2make5]

New working paper quantifying arXiv publication patterns in the age of LLMs! Joint work with @rajivmovva, @sidhikab1, @kennylpeng, @gsagostini, and @NikhGarg.

We analyze LLM citation patterns, fastest growing topics, many other things. Some of our findings: 1/N

[Joelle Pineau @jpineau1]

Llama 2 is out! https://t.co/Tqo1SCtHZs
This new version has better language generation, more layers of safety, a broad set of partners – and a license that authorizes commercial use.
I continue to believe that an open approach is the right path to build better models!

[Dreaming Tulpa 🥓👑 @dreamingtulpa]

Animate-A-Story is a video storytelling approach which can synthesize high-quality, structured, and character driven videos. Composition and scene transitions are still early days, but interesting to see how a first text-to-story pipeline looks like.

https://t.co/XdhYgyzetG

[Matt Sims @mbwsims]

@davecraige for sure

[Matt Sims @mbwsims]

@RoyPrice For sure - if you send me a DM with your email I'll add you to the closed beta.

[james yu @jamesjyu]

We released a big update to the @sudowrite Summarize function - it now generates loglines along with a more coherent summary of the plot

Here's its take on my short story In the Space of Twelve Minutes https://t.co/vX4rtIKLip

Reply with a story and I'll run it!

[Matt Sims @mbwsims]

@Ted_Underwood Yeah, one of the authors' arguments is that the strict left-to-right generation of autoregressive LMs is a strong limiting factor for complex controllable generation tasks, a shortcoming that the denoising process of diffusion models effectively avoids