←
Fundamental Research Labs141 posts
Fundamental Research Labs avatar

Fundamental Research Labs

@Fundamental

Building machines with fundamental human qualities Makers of the strongest spreadsheet agent @tryshortcutai

◷ Joined November 2021
nico avatar
⟳ Fundamental Research Labs retweeted
nico@nicochristie·Jul 24

Opus 5 is a shocking step up from Fable 5 and 5.6 Sol for spreadsheets Almost couldn’t believe our evals at first A step function increase in intelligence, while as efficient as Sol Feels like another Opus 4.5 moment for knowledge work

Post attachment
nico avatar
⟳ Fundamental Research Labs retweeted
nico@nicochristie·Jul 13

I challenged the MSFT Excel World Champion to a battle. AI beat humans at Chess, then Go. But those are games We built an agent to surpass humans on the most important app in the history of work Meet Shortcut: The Excel AI Agent Comment SATYA and I'll send you free credits

3:22
Robert Yang avatar
⟳ Fundamental Research Labs retweeted
Robert Yang@GuangyuRobert·Jul 13

@nicochristie just beat the Excel world champion at financial modeling by using @tryshortcutai 18 months ago, we demoed to investors an agent that's 2x SOTA on OSWorld, and "superhuman" on its Spreadsheet category, scoring ~80% when every other agent was <20% At the time our agent was able to create pivot tables and that was considered impressive We have truly gone a long way, super proud of the team🔥

Post attachment

I challenged the MSFT Excel World Champion to a battle. AI beat humans at Chess, then Go. But those are games We built an agent to surpass humans on the most important app in the history of work Meet Shortcut: The Excel AI Agent Comment SATYA and I'll send you free credits

3:22
Adrian Duermael avatar
⟳ Fundamental Research Labs retweeted
Adrian Duermael@aduermael·Jul 10

What a week! The first GPT model worth being served by default in Shortcut. Let's benchmark Grok 4.5 tomorrow! 😁

we evaluated GPT 5.6-Sol against Fable and Opus using the Shortcut production harness on our own internal set of benchmarks, results were very surprising ShortcutBench V1: swiss-army knife spreadsheet tasks ShortcutBench V2: gnarly, tough, long spreadsheet tasks Sol costs half as much as opus and had similar if not higher accuracy. It also took way less turns and was much faster. Historically, we never served GPT models as default because they are significantly worse at formatting (trust-breaking) than Opus class models. This is also no longer true in our taste tests, although fable is still the best at formatting well done @OpenAI

Post attachment
Peter Wang avatar
⟳ Fundamental Research Labs retweeted
Peter Wang@BrainsAndTennis·Jul 10

we evaluated GPT 5.6-Sol against Fable and Opus using the Shortcut production harness on our own internal set of benchmarks, results were very surprising ShortcutBench V1: swiss-army knife spreadsheet tasks ShortcutBench V2: gnarly, tough, long spreadsheet tasks Sol costs half as much as opus and had similar if not higher accuracy. It also took way less turns and was much faster. Historically, we never served GPT models as default because they are significantly worse at formatting (trust-breaking) than Opus class models. This is also no longer true in our taste tests, although fable is still the best at formatting well done @OpenAI

Post attachment
Robert Yang avatar
⟳ Fundamental Research Labs retweeted
Robert Yang@GuangyuRobert·Jun 29

Tianhang finally on twitter. He was a founding member at Qwen and head of RL at 01.ai There's so much discussions about models from the Chinese labs, but very few people on twitter actually worked there and in the US now

Hey guys, I'm leading LLM research for fundamental research labs. I'll be posting more.

nico avatar
⟳ Fundamental Research Labs retweeted
nico@nicochristie·Jun 29

Tianhang is finally on twitter He was formerly a founding team member and head of RL at Qwen, and leads model training here at FRL / shortcut He will be posting more on twitter now Probably the most cracked person i've ever met. Excited for you to see his work soon!

Post attachment

Hey guys, I'm leading LLM research for fundamental research labs. I'll be posting more.

tianhang zhu avatar
⟳ Fundamental Research Labs retweeted
tianhang zhu@TianhangZhuzth·Jun 29

Hey guys, I'm leading LLM research for fundamental research labs. I'll be posting more.

nico avatar
⟳ Fundamental Research Labs retweeted
nico@nicochristie·Jun 17

Shortcut can now spin up 1000's of agents to parallel search, but we built the feature entirely around the filesystem - enabling some pretty cool new things Here I fill out a huge table of VCs we tracked for a fundraise and another table that has subtables, merged cells, etc.

2:08

Peter Wang @BrainsAndTennis Jun 17 Article Filesystem-Pilling Your Vertical Agent "Filesystem and bash is all you need" has been the de facto way to build agents for at least six months now. Longer, if you date it to the mainstream arrival of Manus. And yet, concrete case studies

nico avatar
⟳ Fundamental Research Labs retweeted
nico@nicochristie·Jun 10

/Autonomous is now default on in Shortcut. @BrainsAndTennis built it before Codex released /goal and has gotten pretty awesome reviews. You need runtime observability and verifiable goals for long term autonomy, and our solution to compacting is to never ever do it at all

I think it's safe to say Excel is solved. The sign of an Excel rookie used to be manually using your mouse... now it's doing ANYTHING at all without AI. If you use Excel do yourself a favor and watch all 3 mins. It's a new world. Solving tasks in 20 mins that take teams days.

3:00
Robert Yang avatar
⟳ Fundamental Research Labs retweeted
Robert Yang@GuangyuRobert·Jun 10

Fable 5 is the best model today, but for everyday spreadsheet task, you won't see a huge difference You'll get much more gain with the right harness & environment Codex + Mog is dramatically better than Codex's default spreadsheet skill, and almost as good as the best agents

Post attachment

Mog v0.9 is published! Mog is the best spreadsheet for agents, and available wherever your agents are Use it in Codex, Claude Code, etc. Just tell them to use npm mog-sdk/sdk, that's all Use it with Shortcut agents, Cloudfare workers, etc Edit xlsx in VSCode, Cursor, etc

Post attachment