Gabriel Gordon-Hall

CTO @bloopdotai | Trying to make LLMs behave

[Gabriel Gordon-Hall @ggordonhall]

@ericzakariasson the best one i've found is Zapier's: https://t.co/F0TS9jrMk9

decent diversity of real-world tasks

[Gabriel Gordon-Hall @ggordonhall]

@farsn_ @theo You can open a worktree in Antigravity if that's what you mean.

Documenting this now: https://t.co/wUEWMDG80y

[Gabriel Gordon-Hall @ggordonhall]

@theo https://t.co/D95LMKjXP2

[Gabriel Gordon-Hall @ggordonhall]

@LouisKnightWebb @vibekanban @AnthropicAI @dedaluslabs @elevenlabs @AITinkerers 🔥🔥🔥

[Louis Knight-Webb @tokengobbler]

I'm super excited to share an early preview of Debugger MCP!

We wanted LLMs to debug and understand code like humans, so we're releasing an open-source MCP server that connects any Debug Adapter Protocol (DAP)-compatible debugger to your favourite MCP client.

It works out of the box with Python and Node, with support for other languages dropping soon.

Debuggers allow you to pause code execution, inspect variable values, step through code line-by-line, set breakpoints, evaluate expressions, monitor program flow, and examine the call stack, making it easier to pinpoint and resolve bugs.

[Gabriel Gordon-Hall @ggordonhall]

I hate having to explain my development context to Cursor, e.g.

"I'm trying to speed up this Docker build, so I've added this caching but now my project crashes with this error..."

I want to screen share with a model so that it can literally *see* what I'm trying to do

[Gabriel Gordon-Hall @ggordonhall]

If you prompt LLMs in Rust I highly recommend `rinja`: https://t.co/E59pn9Gj2a

It's a compile-time template rendering engine that let's you harness the power of Rust pattern matching in your prompts:

[Gabriel Gordon-Hall @ggordonhall]

@jacksbridger They're known as characterisation tests: https://t.co/Oi1fbujQQ4

You run the program with some inputs, collect the outputs, and synthesise tests that `assert f(input) == output`.

If after refactoring `f` a test fails, the program behaviour has changed.

[Gabriel Gordon-Hall @ggordonhall]

It now makes sense to religiously strive for 100% test coverage.

With it, AI coding agents can take on huge maintenance burdens with minimal manual review:

- Dependency upgrades
- Code refactors
- Optimisations

Challenging, but will lead to proper automation

[Gabriel Gordon-Hall @ggordonhall]

@jacksbridger Ideally not. That would be extremely tedious. For this to scale we're going to need better automatic test generation. There are lots of ways it could be automated.

[Gabriel Gordon-Hall @ggordonhall]

@probkirby @LouisKnightWebb I think characterisation tests are about to become all the rage, especially if we improve techniques to generate them cheaply

[Louis Knight-Webb @tokengobbler]

2024 is off to a good start, 120 registrations in 12 hours for January's AI Tinkerer hosted by Monzo! https://t.co/tb34gcPRRz

[Louis Knight-Webb @tokengobbler]

Packed room kicking off AI Tinkerer London tonight. Thanks to @Onfido for 🏡 and @mysticdotai for 🍕!!!

[Louis Knight-Webb @tokengobbler]

Tool Augmented Language Models.

The #AI4Code reading group is also back in July with Aaron Parisi, #Google researcher and lead author of the #TALM paper👇

https://t.co/5pwNlZpSLC

[Gabriel Gordon-Hall @ggordonhall]

Can't wait for this!

[Markus Strasser @mkstra]

I spent months working on domain-specific search engines and knowledge discovery apps for biomedicine and eventually figured that synthesizing "insights" or building knowledge graphs by machine-reading the academic literature (papers) is *barely useful* :
https://t.co/eciOg30Odc

[Philip John Gorinski @xylankant]

if you're attending #acl2020nlp why not say hi in our live Q&A sessions where @ggordonhall and myself will answer any and all of your questions about "Learning Dialog Policies from Weak Demonstration" (https://t.co/tAqDzdnWG0)
#huawei #nlp #AI #ReinforcementLearning

[Gabriel Gordon-Hall @ggordonhall]

Can data predict fashion trends? By me @TheEconomist
https://t.co/y7iXSOUD8P