Ashu (@ashu) header
Ashu (@ashu) profile picture

AI/Tech/Management

United States
Joined October 2009
Photo by Ashu (@ashu)Photo by Ashu (@ashu)Photo by Ashu (@ashu)Photo by Ashu (@ashu)Photo by Ashu (@ashu)Photo by Ashu (@ashu)Photo by Ashu (@ashu)Photo by Ashu (@ashu)Photo by Ashu (@ashu)Photo by Ashu (@ashu)
300+ humanoid robots ran a half-marathon in Beijing alongside humans and broke the human record. We went from “robots can walk” to “robots can race and beat humans” in about 18 months. Pace of progress is not slowing down. #Robotics #AI
415
Cloudflare just released a lossless LLM compression technique that cuts weight size without losing model quality. Turns out BF16 exponent fields in trained weights only carry ~2.6 bits of real entropy despite using 8 bits. That's a lot of wasted silicon, and this closes the gap.
404
Stakeholders don’t derail projects because of technical problems. They derail them because of information vacuums. Fill the vacuum first.
291
Precision wins. CMS just measured the W boson mass and it lines up with the Standard Model after years of debate. Huge data, sharp simulation, and relentless analysis = the kind of science that keeps pushing the frontier forward! nature.com/articles/s41586-0…
231
#Anthropic had the most powerful hacking tool ever built - and chose not to ship it. Instead: $100M in credits to defenders, early access to Apple, Google, Microsoft to patch critical systems first. This is what responsible AI development actually looks like in practice. 👏
162
Anthropic’s Mythos model is reportedly being limited before release because it’s too good at hacking. Let that sink in. We’ve built something so capable at finding vulnerabilities that the builder itself is afraid to ship it. This is the dual-use dilemma of our generation.#AIRisk
130
My preordered copy of Planet Money book arrived this morning. Looking forward to reading it over the weekend! planetmoneybook.com/
102
One habit that pays off in large‑scale ML: instrument everything. Log gradients, activations, data stats, throughput, and failure modes early. Half of ‘model magic’ is just having enough visibility to know why something improved. Observability turns intuition into engineering!
88
Your roadmap is not a plan. It’s a hypothesis. The teams shipping the best products treat every quarter like a scientific experiment - with expected outcomes, leading indicators, and kill criteria.
79
The real AI bottleneck isn't the model; it's the I/O. Scalable inference is a systems engineering problem. If you aren't optimizing memory bandwidth and kernel execution, you're just burning money on idle GPU cycles.
82
Are we focused on winning the wrong AI race?: platforms.substack.com/p/us-…
316
In the meantime there’s SAS Rogue Heroes - S2 for watching
1
1,458
And “the Recruit” is only a week after!
1,430
Is anyone else excited about the new season of “the night agent”?
1,383
Past actions can affect your future, but what if future actions can affect the past (and present)? “Forward to the Past” anyone? https://teapawt.net/t.co/SnSvSwVNDi
1
2
2,451
Join me on #PiDay as we demo AI/ML/NLP capabilities of #OCI A1 #arm compute instances. Register early for free swag!
A tech event by devs, for devs. This #PiDay, join other #developers and your favorite streamers for a celebration of Pi and what’s possible with our #cloud. social.ora.cl/6015Kk3GJ:
2
1
It’s router patching time! https://teapawt.net/t.co/xoBXgRXh8X
1
Help stop the spread of COVID-19 by using an exposure notification app for your state/region/country. Find your app here: https://teapawt.net/t.co/DxsqsudXB3
1