Mo

197 posts

Mo banner
Mo

Mo

@goesbymo

AI Lead @ ABX

Brooklyn, NY Katılım Temmuz 2024
110 Takip Edilen145 Takipçiler
Sabitlenmiş Tweet
Mo
Mo@goesbymo·
Conservation of Complexity: Ct = Cb + Cu The system you build will always have a constant complexity, Ct, that is equal to the building complexity (Cb) + the UX complexity (Cu). For example, you'd like to build an app that files people's taxes. If you want to minimize Cu, and do something like a single button that a user clicks on to file their taxes, Cb is going to go through the roof because you'd need to make a lot of assumptions, pull their W-2s or 1099s from somewhere, check all brokerage platforms for whether the user has sold any stock, etc. Minimizing Cb would look like simply accepting files and parsing them for details, a vibe codeable app honestly, but the UI will need to contain a large form forcing the user to find and upload every single document that's needed. when you build, think of this formula and where you'd rather shift the complexity. this framework is fractal, and can be applied to everything from an individual feature to an entire product. I personally used it a lot in the VC backed days, and it did wonders for my customers.
English
2
0
12
971
Mo
Mo@goesbymo·
Grok 4.5 is my new daily driver. might be the first time both my FPM and API usage on cursor all goes to one model.
English
0
0
1
53
Mo
Mo@goesbymo·
cyka blyat
Mo tweet media
Tim Lupo@timlupo

Excited to share that @deeptuneai is joining @mercor_ai. When we started Deeptune, we believed RL environments would become the defining bottleneck for AI. Over the past two years, we've seen this play out even faster than we expected. Mercor has built the world's leading expert network for AI, making it the obvious place to continue our mission. Together, we'll build the platform that powers the next generation of AI agents. Our entire team is joining Mercor in NYC, and we'll continue building with far greater resources towards the even bigger opportunity ahead. I'm deeply grateful to our customers, investors, and everyone who helped us get here. Most of all, thank you to the Deeptune team. None of this would have been possible without you. We're just getting started.

Svenska
1
0
2
1.3K
Mo
Mo@goesbymo·
@jerryjliu0 btw, KDL-frontier-Parser-nano was trained on the parsebench evaluation set, and didn't perform well at all with real world documents I handed it. is there a disclaimer/note you can add to models like that on the leaderboard?
English
0
0
0
36
Jerry Liu
Jerry Liu@jerryjliu0·
We comprehensively benchmarked GPT-5.6 on document understanding. At a high-level there's no change between GPT-5.6 Sol and GPT-5.5 in terms of performance over tables, text, charts, layout, and more. The GPT-class of models typically does quite well over table understanding, but they struggle with transcribing complex text layouts and formatting, with transcribing charts, and with deriving bounding boxes over source elements. Come check out our leaderboard of over 70+ frontier models, open-weight models, and OCR solutions on ParseBench: parsebench.ai
Jerry Liu tweet media
LlamaIndex 🦙@llama_index

@OpenAI released GPT 5.6 today and we ran a day 0 benchmark in ParseBench to test improvements in document understanding. The new family of models continues to excel at reading text and tables, but continues to struggle with charts and layout. What's most interesting is that Luna is about 6 x cheaper than Sol and only results in minor degradations across all ParseBench metrics, showing that an increase in reasoning tokens doesn't always result in a commensurate improvement in visual understanding.

English
12
11
115
18.9K
NoLimit
NoLimit@NoLimitGains·
Jim Cramer on Nvidia $NVDA: “I think you should own Nvidia and not trade it”
English
159
32
533
421.4K
Mo
Mo@goesbymo·
I have a feeling that this is going to be the undoing of Anthropic. we're still a ways away from that given how prominent they are in the coding world, but other labs are catching up at a fraction of the cost. perhaps soon, anything over $50/Mtok output would be obscene.
Vals AI@ValsAI

Meta just released Muse Spark 1.1 and is the new SOTA on MedScribe and TaxEval, taking the top spot from Fable 5 while being 10x cheaper and twice as fast. Meta currently holds the top 2 spots on TaxEval It is also the new #1 on Harvey's Legal Agent Bench, dethroning Grok 4.5 less than 24 hours after it took the top spot.

English
0
0
0
50
Mo
Mo@goesbymo·
@jordankdalton indeed, pretty insane how much time (and tokens) a little bit of planning saves
English
0
0
1
17
Jordan Dalton
Jordan Dalton@jordankdalton·
@goesbymo Planning I feel is the key skill now. Too many hurry to ship where just a little of planning would save them time and quality.
English
1
0
1
12
Jordan Dalton
Jordan Dalton@jordankdalton·
When’s the last time you wrote code without AI?
English
86
0
21
8.7K
Mo
Mo@goesbymo·
Pre-AI I had the skills of a junior engineer, but using them has taught me a lot about both planning and writing software. At work where there’s deadlines and shareholders, I’ll use AI. But programming is still fun, so when the pressure’s off I like to make use of my fancy keyboard
English
1
0
1
51
Mo retweetledi
Tibor (Tee)
Tibor (Tee)@tibor_tee·
Grok 4.5, Composer and Auto pool is now our first party models pool.
Mo@goesbymo

@tibor_tee you legend❤️

English
20
7
144
9.3K
Mo
Mo@goesbymo·
@bizdevelo Charles was 74 at his coronation, you still got time king👑
English
1
0
1
16
Yas
Yas@bizdevelo·
@goesbymo I’m too late to the party then
English
1
0
1
98
Yas
Yas@bizdevelo·
unpopular opinion: founders 30+ y/o perform much better, work more efficiently and have a higher success rate so why tf I gotta be 19 to get into YC again???
English
44
3
114
13.1K
Mo
Mo@goesbymo·
@amasad when paradigm changes as rapid as compilers to hand-writing assembly happen, the benchmarks are always lousy in the first few years. I feel like they'll start making sense when agents are reliably compared to older agents.
English
0
0
0
97
Mo
Mo@goesbymo·
@benln mentioned this in the cursor post, but please move its usage under Auto + Composer. their capabilities with your real time RL would make it so Fable doesn't keep taking 3/4 benchmarks in this image.
English
0
0
1
67
Mo
Mo@goesbymo·
@cursor_ai big proponent of sharing this model's usage with Auto + Composer, as many replies here have voiced. Grok's capabilities with your real-time RL🔥
English
0
0
1
85
Cursor
Cursor@cursor_ai·
We've partnered with SpaceXAI to train Grok 4.5. It’s our most powerful model yet and the first we've built for more than software engineering.
Cursor tweet media
English
729
1.5K
17.7K
5M
Mo
Mo@goesbymo·
only way I'll actually use Grok 4.5
Mo tweet media
English
0
0
4
198