A pipeline for creating VQA data using knowledge bases (wikidata). We demonstrated the method on world/cultural knowledge and could be extended to other domains. (there are rephrasing/filtration steps that require LLMs/VLMs, but the released data does not).
@jeande_d
-
Andrej Karpathy’s 2018 talk on Building the Software 2.0 Stack
By
–
The talk is "Building the Software 2 0 Stack (Andrej Karpathy)" 2018
-
Karpathy: PhD sleep lost over models, industry over datasets
By
–
There is this 2018 talk of Andrej Karpathy comparing his PhD time spent vs Tesla old days. PhD: sleep lost over massaging models and algorithms
Tesla(industry): sleep lost over massaging datasets In reality it is always about data and evals/your mechanism of evaluating the -
Flywheel of models and data limited to verifiable domains
By
–
Models, synthetic data, and environments are a flywheel now. The best model generates the best data, trains a better model, generates even better data. Rinse and repeat. But the flywheel only spins where you can verify: math, code, agents with graders. Everywhere else it merely
-
VQAScore: open-source framework for evaluating text-video/image/3D models
By
–
VQAScore, a simple framework for evaluating text-video/image/3d models is opensource and got many many features. Packaged with leading frontier and commonly used open-source multimodal models and simple/clean eval interface. https://t.co/8xy7SjGEqt pic.twitter.com/wrt6wBTQvQ
— Jean de Nyandwi (@Jeande_d) 9 juin 2026VQAScore, a simple framework for evaluating text-video/image/3d models is opensource and got many many features. Packaged with leading frontier and commonly used open-source multimodal models and simple/clean eval interface.
-
Karpathy’s 2015 RNN coding experiments and 2017 gradient descent advances
By
–
2015: Karpathy publishes one of the earliest experiments (if not the first public article/blog with practical artifacts) on language models for coding (The Unreasonable Effectiveness of Recurrent Neural Networks). 2017: @karpathy again: gradient descent can write code better
-
Personal LLM Knowledge Bases: The Most Effective Learning Method
By
–
Building and maintaining personal LLM knowledge bases might be the most effective way of learning things with LLMs. Consuming/summarizing things off the chat box does not really seem to work.
-
Multi-Agent Software Engineering: Branch-and-Merge Coordination Strategies
By
–
As long-horizon software engineering tasks grow in complexity, a single agent can no longer finish the tasks alone — effective multi-agent collaboration becomes necessary. This leads to a natural question: how can multiple agents be coordinated to asynchronously collaborate over a shared artifact in an effective way? We answer this question in our new preprint: Effective Strategies for Asynchronous Software Engineering Agents! We suggest that to coordinate multiple software engineering agents, branch-and-merge is the key coordination mechanism, and that human SWE primitives like git worktree, git commit, and git merge are all you need to support it. (1/n)
-
New LM Reasoning Benchmarks for Complex Mathematical Objects
By
–
🧮New work from @AIatMeta & @LTIatCMU! LM reasoning benchmarks mostly use simple answers like numbers (AIME) or multiple-choice options (GPQA). But for complex mathematical objects, performance drops sharply. We propose a set of solutions to solve this: arxiv.org/abs/2603.18886
-
AI Agents Benchmarks Don’t Match Real Human Work
By
–
AI agents are tackling more and more "human work" But are they benchmarked on the work people actually do? tl;dr: Not really Most benchmarks focus on math & coding, while most human labor and capital lie elsewhere. 📒 We built a database linking agent benchmarks & real-world work Submit new tasks + agent trajectories today 🧵