I’m rooting for you. raise AGI with the richness of humanity.
Austin, TX
Joined July 2010
- Tweets629
- Following1.3K
- Followers767
- Likes4.6K
A lab releases a new model, scores impressively on major benchmarks, "is this AGI??!!", but they still can't understand PDF's.
Enormous upside remains for real-world utility and for enterprises to make the most of these tools.
dave games retweeted
🎤 Who run the world? 🎤
Gir—PDFs. PDFs run the world.
This week, we launched GDP.pdf: a new, expert multimodal reasoning benchmark.
We've spent years measuring AI against the extraordinary: proving theorems, solving AGI.
But the global economy doesn't run on the extraordinary.
It runs on paperwork.
More precisely: unsexy, poorly scanned, densely formatted PDFs. Contracts, invoices, medical records, blueprints – the documents that underlie everything we do in the enterprise.
So GDP.pdf tests frontier models on their ability to handle real-world documents across ten professional industries:
🏗️ Construction: Can a model measure load-bearing walls on a blueprint?
⚖️ Law: Can it parse liability caps in a commercial lease?
💵 Finance: Can it calculate margin profiles in a buy-side memo?
Every frontier model scored under 15%.
With GDP.pdf, we wanted to ask: if a $100B model can’t accurately reason about a drug interaction table in a PDF, is it actually ready to take over the economy?
Right now, the answer is no.
Check out the blog post and leaderboard below!
Blog: surgehq.ai/blog/gdp-pdf-can-…
Leaderboard: surgehq.ai/leaderboards/gdp-…
He was right about the metaverse, but wrong that it would be for humans talking to humans.
Meta builds AI version of Mark Zuckerberg to interact with staff ft.trib.al/9Nx1dvW
If you work in tech and are also a prepper, people really read into your actions as a signal for the end times.
I am by no means dictating any AI trends, but I am the resident 'AI tech member of the family', akin to your grandparents asking you to fix their tv because you 'work with computers'. I gifted a few members of my family some spear tips, and of all things, I was not expecting their reaction to prompt conversations about AI safety.
I can’t explain it but Russian sounds actually make more sense coming from ravens. Like they spoke it first.
badass
A new product, a new customer, a new financing!
Introducing Superpower: a 42MW natural gas turbine optimized for AI datacenters, built on our supersonic technology. Superpower launches with a 1.21GW order from @CrusoeAI Backstory 🧵👇
It's interesting to see how different models perform in the core agentic skill sets.
"Models that can’t do this reliably aren’t agents; they’re chatbots with access to tools."
Everyone's acting like models are ready to replace humans in work settings.
We put that to the test by creating an entire company and having 9 models act as a customer service agent handling 150 tickets and requests of increasing complexity.
Verdict: without common sense, models are nowhere near ready.
👇
surgehq.ai/blog/rl-envs-real…
The name. The timing. No notes.
A NEW PAPER FROM YANN LECUN:
LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics
This could be one of LeCun's last papers at Meta (lol), but it's a really interesting one I think. Quick summary:
Yann LeCun's big idea is JEPA, a self-supervised learning method. However, there are various failure modes of this approach, so training strong JEPA models is very brittle, unstable, and quite difficult. So overall JEPA has seen little adoption in practice.
This paper tries to directly address this, making specific design decisions that improve training stability.
The authors identify the isotropic Gaussian as the optimal distribution that JEPA models’ embeddings should follow and design the Sketched Isotropic Gaussian Regularization (SICReg) to constrain embeddings to reach that ideal distribution. This forms the LeJEPA framework, which can be implemented in ~50 lines of code.
On empirical tests, the authors demonstrate stability of training across hyperparameters, architectures, and datasets.
A result particularly interesting to me however is that training a LeJEPA model from scratch directly on the downstream dataset outperforms finetuning a DINOv2/v3 model on the dataset!
People knock this as a destination for tourists, and they should. It’s an okay way to see part of the city, but it’s a great way to commute on the west side. Walking through historic garment warehouses into the new Hudson Yards is energizing and connecting to the city. It’s peaceful enough to actually consider where you are.
Careful. You’re about to change your view of the South with the nicest beach towns in the country.
This quoted post is unavailable.
Especially in today’s enterprise tech environment that leans more technical with API’s, embedded AI tools, etc, most sales teams do not have the core tech strength they will need. Look at solution architect or solution engineer roles if pure sales is not the path you want, you’ll also stay closer to eng that way. Careful though, you might enjoy being close to the customer problems.
dave games retweeted
I have so many thoughts on this, as someone who has run large EPD orgs.
One of the toughest parts about doing an AI operating transformation inside a prod/eng/design org is that people start operating out of their lane and it stresses SO MANY PEOPLE out.
PMs can vibe code prototypes -> eng is stuck with aspiration + code that aren't production ready
Designers can add in more polish and craft -> scope starts to creep
Engineers can vibe-code new features -> who asked you to build that
But the reality is: this isn't really an AI problem. The inefficiencies of of "stay in your lane," "this is too complicated," "i own the prioritization" have always existed, they've just been codified into the operating manual of most orgs so they feel right ("it's how things are done") and they make everyone on the team feel happy, because it maximizes individual loci of control (PM owns this, design decides that, you get whatever eng decides to build) vs the velocity of the system.
AI starts to crack the facade of these fiefdoms, in more obvious and frequent ways, and you start to experience dumb conversations like below often, with a stronger sense that there's a better way to build.
But humans cling to their old ways! And the operating model you've built is calcified in such a way, it's easier to give up and stay in your lane vs. push for a new model.
But I have faith. I think if you can change your *culture* you can change how you build, for the better.
Repeat after me:
- make experiments cheap: try that fancy design coded w AI, you may get further than you think. or you can throw it out without too much waste
- get good with AI tools: you can marry a vibe-coded demo with a production-ready codebase + AI IDE and actually get pretty far with quality code
- there are no lanes: welcome people coding, designing, PMing even if it's not their official title
- make it fun: your roadmap is not life or death. the more people are having fun, vibing (literally), and don't feel the crushing weight of an organization-level anxiety disorder, the faster they can go
Finally, you can just build things!
(And be nice to your colleagues, please.)
I like grocery shopping. I like coming across an interesting cheese or fruit that my kids can try, or seeing a steak or fish that dictates a new dinner plan tonight.