Login or sign up
For business
Average ratingAverage
3.6/ 5

Based on 5 reviews

540%
420%
320%
20%
120%
Ask AI

Questions about david kogan

Based on 5 reviews, listing data and website checks · updated 5 days ago

What is david kogan?

David Kogan is a self-taught AI researcher and the founder of an open source AI efficiency lab focused on making frontier AI cheaper to train and run. He built his first PC at 11 and maintains a personal website describing his background and projects. He currently serves as a technical advisor to duino.

What specific tools or projects does the lab offer?

The lab focuses on upstream projects like hypertok, which is the fastest tokenization on the web, achieving up to 295x speed improvements over js-tiktoken. Other tools include turboadam, a drop-in optimizer replacement that is 9.845x smaller than adamw, and dcaf/noosphere interpretability tools for circuit discovery.

What are the key technical innovations being developed?

The website states that the lab is building an ffn that skips about 85% of per-token ffn compute, performing within about 6% of standard in small-scale tests. They also research a low-bit residual master weight at about 89% perplexity parity and a ternary training format.

Where is the business located and what is its background?

The lab is located in Brooklyn, New York, and the founder previously attended Technion pre-engineering before leaving after a missile hit the campus, then studied economics at Binghamton.

What is the core mission and operational style of the lab?

The lab aims to make frontier AI radically cheaper to train and run, with a goal that science shouldn't be this expensive. It operates as a conceptual builder where the founder designs and agents work, rather than following traditional coding structures.

What is the current rating and review count?

The Trustburn rating is 3.6 out of 5 from 5 reviews. The listing includes a bio highlighting his pivot from hardware-enforced alignment to open-source efficiency, noting that safety is not control.

What practical solutions have been provided for developers?

The lab has produced prebuilt CUDA wheels called easywheels to avoid 30 to 120 minute source builds, and retired when uv's team started building the same thing.

What attention mechanisms were designed and tested?

The founder designed and tested entropy-attention and r-attention mechanisms, though neither beat standard attention across an 8-task battery. Despite this, both were published as part of the research output.

What safety and testing processes have been applied?

The lab has undergone many rounds of AI red teaming, though no human expert has checked the specific outputs mentioned. The founder also spent six months on hardware-enforced alignment before pivoting.

What is the official website of david kogan?

The official website of david kogan is davidkogan.org.

AI-generated from reviews on this page. Verify anything important with the company directly.

Sarah Jenkins
Reviews 1

The ease of accessing these tools is a major plus. Having hypertok available on npm and turboadam on pypi means you can just pip install or npm install and get to work immediately. The detailed benchmarks help you understand exactly what you are getting, which saves time on evaluation. The fact that easywheels provided prebuilt cuda wheels is also convenient, even if it has been retired. For anyone building with ai models, having these specific optimizations readily available online makes the development process much smoother and less dependent on complex build environments.

Robert Sterling
Reviews 1

While the technical specifications for tools like hypertok are detailed, the website does not explicitly state pricing or licensing terms for commercial use. It is unclear whether the open source nature applies to all components or if there are hidden costs for enterprise support. Potential clients should verify the scope of the open source license before integrating these tools into critical infrastructure, as the current information leaves room for ambiguity regarding commercial availability.

Elena Vasquez
Reviews 1

It is refreshing to see an open source ai efficiency lab with such clear project repositories. The availability of hypertok on npm and turboadam on pypi makes integration straightforward for developers who need faster tokenization or smaller optimizer states. The detailed benchmarks provide transparency, allowing users to verify the claimed speedups before adoption. The presence of exact token ids in the browser and node environments suggests a strong focus on practical usability. This approach reduces the barrier to entry for those who want to test the technology in their own workflows without needing deep expertise in the underlying architecture.

Marcus Thorne
Reviews 1

The technical details presented on the site are impressive, particularly the specific performance metrics for hypertok and turboadam. However, the narrative is dense with jargon like enstrophy-weighted directional fisher information. It reads more like a research log than a service catalog. For those who understand the underlying mathematics, the efficiency claims are compelling, but the average user might struggle to grasp the practical value of these innovations.

News

No news yet. This is where the business shares its own updates and press.

Photos

No photos yet. Photos shared by reviewers will show up here.

Widgets

Boost your business trust by displaying a widget on your website — a simple way to convert your visitors into customers.

Get the widget for your site →

About david kogan

David Kogan is a self-taught AI researcher who maintains a personal website describing his background and projects. He is building an open source AI efficiency lab focused on making frontier AI cheaper to train and run, and has produced upstream projects including a fast tokenizer, an optimizer replacement, interpretability tools, prebuilt CUDA wheels, and research on low-bit training methods. He also serves as a technical advisor to duino.