Phuc

post-training & infra @nousresearch | ex-nanotron at @huggingface | dms open

[Phuc @phucnhx]

it was fun to work on this. check out the blog post if you want to see how to get expert parallelism to scale linearly through internal worklogs and other kernel-level optimizations!

[Nous Research @NousResearch]

Follow along with researcher @xariusrke in these detailed field notes as he hunts down MoE training bottlenecks

https://t.co/0bVb2Sb0bf

[Polymarket @Polymarket]

JUST IN: Elon Musk reveals future Roadster demonstrations could include flying over the Grand Canyon.

[Phuc @phucnhx]

@NousResearch @Teknium The goat of all goats

[Tesla @Tesla]

Real life, baby

[dmayhem93 @dmayhem93]

Hermes Bench mentioned 👀

[Phuc @phucnhx]

@Tu7uruu @reach_vb yes, sometimes they reset it for all users https://t.co/wL3OMFEJfi

[Phuc @phucnhx]

@iamJonasB @NousResearch okay cool lmao

[Phuc @phucnhx]

@iamJonasB @NousResearch from fashion people?

[Phuc @phucnhx]

@fl_rn_st @NousResearch Welcome!

[Vega Shah @dr_alphalyrae]

Building on decades of joint work with national labs, NVIDIA has committed $1B towards developing advancements in US science. For research across quantum computing, healthcare and energy.

Many years ago, as a postdoc at lawrence berkeley I saw how industry partners were integral to applied research at the federal level and also critical to training the scientists and engineers of the future

[Phuc @phucnhx]

@Tu7uruu

[Phuc @phucnhx]

> Step 5 Preview scored 33.89 on Hermes Index, the same score as GPT-6 Luna

[Nous Research @NousResearch]

Step 5 Preview from @StepFun_ai is free on Nous Portal for the next week

600B total / 27B active MoE, 1M context, with vision

Step 5 Preview scored 33.89 on Hermes Index, the same score as GPT-6 Luna

[Yuchen Jin @Yuchenj_UW]

Saw a video yesterday:

A math PhD and her advisor spent 6 months proving a theorem. Before publishing, they asked ChatGPT.

AI found a more elegant proof quickly, using an approach they'd never considered.

They have an existential crisis now.

But mathematicians shouldn't despair. We went through this in coding. Now we can't live without AI.

Deep domain expertise + knowing how to leverage AI is a strong moat in this era.

[steven @Tu7uruu]

Just dropped the URGENT 2024 dataset on the Open ASR Leaderboard!

URGENT 2024 benchmarks speech enhancement under real-world distortions: noise, reverberation, clipping, and bandwidth limitations.

How robust are ASR models to reverberation? 🎙️

Most models see their WER roughly double!

→ Zoom: just +41% WER
→ Cohere: lowest noisy WER among open models
→ IBM Granite: smallest WER increase among open models

[Kalshi @Kalshi]

JUST IN: Trump says Elon Musk is the "modern-day Thomas Edison"

[Phuc @phucnhx]

@qorprate no the singular is near? and sapiens?

[Cowboy Space Corp. @CowboySpaceCorp]

We’re convinced that larger satellites are the best foundation for orbital AI compute. That conviction drives our design at Cowboy Space.

AI workloads need GPUs to exchange data quickly, including the working context used during inference. Larger satellites can host larger clusters of GPUs and keep more of those exchanges on onboard networks, keeping latency to a minimum and reducing reliance on inter-satellite links for tightly coupled workloads.

At Cowboy Space, our approach is one large satellite per launch with a megawatt of peak compute capacity. The data center, launch vehicle and operations are designed together to target high bandwidth, low latency and best-in-class mass per unit of compute.

As the power density of compute systems grows, our megawatt satellite can keep up. With fewer satellites required for a given amount of compute, we reduce orbital congestion and improve space safety relative to more distributed constellations.

We’re revealing our roadmap and design at the end of this month. Stay tuned.

[Phuc @phucnhx]

@ethernet_8023 we have a dj at home @gigant_theo

[Watcher.Guru @WatcherGuru]

JUST IN: 🇺🇸 President Trump to award National Medal of Science, the highest honor for scientific achievement to the following tech leaders:

• Lisa Su, $AMD
• Elon Musk, $SPCX
• Sergey Brin, $GOOGL
• Jensen Huang, $NVDA

[Phuc @phucnhx]

cool stories

[witcheer @witcheer]

we just added 45 new stories to this page, all from people using Hermes Agent for work, such as:

- Razorpay gives 220+ employees their own always-on agent
- a pool service company flags bad technician reports while the route is still live
- a physician saves 30 to 45 minutes of documentation per case with Hermes skills
- a Swiss web agency drafts its client invoices, and a human sends them
- an attorney who also owns a local business automates its payment reconciliation
- a bookstore runs one agent for its staff and one for its customers

371 real stories now, and plenty more business ones in there

[Phuc @phucnhx]

@witcheer oui

[Phuc @phucnhx]

> approximately 2.5% of all global token usage

[Nous Research @NousResearch]

As reported in the @WSJ, we have raised a Series B to bring Hermes Agent to new frontiers (and, yes, build a mobile app)

Thank you to our investors including @nvidia @M12vc @SamsungNext @robotventures @usv @ycombinator @MenloVentures, among others

https://t.co/1BdQjJD90p