AT

- 230 training runs - 1,623 GPU hours (67 B200 days) - 76 TB of training data - a 2x faster model Every paper said it can't be done. Quantization Aware Distillation made it possible.

@amiruci product-mogging the competition

We now have a product specifically created for AI labs and their closed-weight models: we'll take care of not just inference, but auth, rate limits, metering, and billing integrations. We'll take care of providing both shared and dedicated inference, compliance needs, and matching end customers' geo requirements (us, ca, eu, uk, aus, jp, etc). It's called Baseten Frontier Gateway and is already battle-tested by multiple AI labs, like Poolside and their impressive Laguna M.1 agentic coding model.

@modal this is a sick read...hats off to you guys

@philipkiely HAHAHAHAHA

Developing empathy for LLMs by doing benchmark problems by hand.

this was so fun to work on, i hope you find it useful tried @baseten for GPU access?

@edenchan solid to the power of solid squared

Think fast! This is the voice that powers customer support and sales for Starlink

@Millanphilipose @kanjun your own WHAT... flex 🙈

@AliesTaha @kanjun We're using our own datacenter for now

Twitter’s algorithm is optimized for addiction, not for us. We deserve better. We’re releasing Bouncer today so you can take back control of your feed. Describe what you don't want, and Bouncer removes it. It’s free, doesn’t collect your data, and will be open source soon.

@anandcpatelmdms @part_harry_ x.com/Goosewin/statu…

@AliesTaha @part_harry_ Smelly?

we dug into 1-bit bonsai with @part_harry_ the grand canyon of a gap they showed... is just THREE (3) points away from normal PTQ but they already knew that here's the graph (fixed)

@nisten @part_harry_ we used their axis to plot on their chart their benchmarks to get the intelligence scores the x-axis is the weight file size this is what PrismML used

@AliesTaha @part_harry_ The graph compares model sizes not total memory use dumbass. You're comparing total kv of 1bit float16 vs finetuned fp4PTQ /fp8 kv at 4k context benchmark or like... what are you even comparing? x.com/nisten/status/…

@HenkPoley @part_harry_ fair, the point is more that the graph was designed to make 3 points look like a generational leap

@AliesTaha @part_harry_ 3 percentage point better is still quite a bit better. 🤷‍♂️ 73.8 to 76.8, about 11% less errors on these tests. Given that most of these tests have errors, so a perfect score cannot be achieved, probably even a bit better.

@JoshPurtell @part_harry_ more eyes on benchmarks is only a good thing

@AliesTaha @part_harry_ Taking this as permission to publicly sanity test forthcoming Baseten results/research

@oneill_c whoaaaaa

on-policy for the student off-policy for the teacher monkey input, monkey output

@philipkiely what is inference? how does it work? @philipkiely can i come to learn (and also maybe get ice-cream)?

Ice cream and books were a hit yesterday. ICYMI we're doing another, this time at the Ferry Building. Thursday 4/2 from 2-4 PM: luma.com/khxc93ju

@gaoj0017 only 3. Their experiments used single-core CPU for RaBitQ vs A100 GPU for TurboQuant has merit as a complaint the other 2 just don't hold

We need to publicly clarify serious issues in Google’s ICLR 2026 paper TurboQuant. TurboQuant misrepresents RaBitQ in three ways: 1. Avoids acknowledging key methodological similarity (JL transform) 2. Calls our theory “suboptimal” with no evidence 3. Reports results under unfair experimental settings We have expressed our concerns to the authors before their submission, but they chose not to fix them in their paper submission. The paper was accepted at ICLR 2026 and heavily promoted by Google (tens of millions of views). At that scale, uncorrected claims quickly become “consensus.” Facts: 1. RaBitQ already proves asymptotic optimality (FOCS’17 bound) 2. TurboQuant uses the same random rotation step but misses stating the connection 3. Their experiments used single-core CPU for RaBitQ vs A100 GPU for TurboQuant None of these is properly disclosed. We’ve filed a formal complaint and posted on OpenReview (openreview.net/forum?id=tO3AS…). We’ll release a detailed technical report on arXiv. Our goal is simple: keep the academic record accurate. Would appreciate people taking a look and sharing.

The TurboQuant paper (ICLR 2026) contains serious issues in how it describes RaBitQ, including incorrect technical claims and misleading theory/experiment comparisons. We flagged these issues to the authors before submission. They acknowledged them, but chose not to fix them. The paper was later accepted and widely promoted by Google, reaching tens of millions of views. We’re speaking up now because once a misleading narrative spreads, it becomes much harder to correct. We’ve written a public comment on openreview (openreview.net/forum?id=tO3AS…). We would greatly appreciate your attention and help in sharing it.

@Phenomenon_One well both really accesible via simplifying it confirm their claims with the gpu kernels, and my verdict was not useful at current perf for latest gpus

@AliesTaha Update after perusing I see where the time went deep diving algos and testing claims on arbitrary datasets .. are you trying to make it accessible or are you trying to confirm their claims?

@Phenomenon_One i think partly due to my background being rooted in swe and not maths, but this was total time, including time it took to write article and make graphs

@AliesTaha Lmao 31 hours 😂 How long did the original researchers spend on it … Will critique after I read this.

@no_identd @windsheep_ looking into it now

@AliesTaha @windsheep_ Should factor in x.com/i/status/20375…