Head of Models @Cloudflare. Founder at @jamdotdev.
Austin, TX
Joined December 2011
- Tweets3.5K
- Following899
- Followers1.6K
- Likes14.3K
Irtefa retweeted
Cloudflare’s Matthew Prince brilliant thoughts on neoclouds in their most recent 3Q26 earnings call:
”We are an engineering company. We are not a financial engineering company...
We are focused on how we can actually deliver real operational excellence…
You can’t sell $1 for $0.50 and be in business over the long term.”
Irtefa retweeted
Google's new video model from @GoogleDeepMind
- 15-second clips, up from 10
- More natural dialogue and lip-sync
- Characters stay consistent across shots
- Edit the same clip with a new prompt
Irtefa retweeted
2 | @Cloudflare
Every Spare Gigabyte: A Fleet-Wide Elastic Mooncake Cache
Kastan Day on turning spare RAM and SSD into a shared KV cache that grows and shrinks with model demand.
Irtefa retweeted
A hop-by-hop view of @Cloudflare network engineer @tstrickx crossing the Atlantic on Starlink Wi-Fi.
Over 8 hours and 7,000 km, Starlink handed off satellites every 15 seconds and carried traffic through 7 Cloudflare deployments. Each plane appears to keep its own IPv6 /56.
Over the ocean, inter-satellite lasers carried traffic to the nearest available ground station. Cloudflare’s anycast routing followed it, keeping traffic local as the plane moved at 40,000 feet and 500 knots.
- Ocean crossings use inter-satellite laser links until traffic can reach a ground station.
- Those laser routes appeared slower than fiber in this single test.
- Latency showed two distinct bands, possibly from first-packet dataplane setup.
- Ground-station selection changed during the flight, including a Canada-to-US switch near the border.
- Cloudflare’s anycast routing followed the ground station, not the plane’s source address. That kept traffic local without DNS-based steering.
starlink.ecmp.network/f/2026…
Irtefa retweeted
Building an iOS or MacOS app in 2026 and have no idea what you're doing? Point your agent at the GitHub profiles of @fayazara and @mattt for reference and you'll be fine.
github.com/fayazara
github.com/mattt
50% off on Replicate and @CloudflareDev until the end of the year!
Thanks @MiniMax_AI !
Minimax H3 Lite and Highspeed are now 50% off until the end of the year.
Exclusively on Replicate and @Cloudflare AI Gateway.
@MiniMax_AI
Irtefa retweeted
Minimax H3 Lite and Highspeed are now 50% off until the end of the year.
Exclusively on Replicate and @Cloudflare AI Gateway.
@MiniMax_AI
Irtefa retweeted
a couple of hours after getting an initial denial of his residency permit, @deloreyjmi is still trying to hire more engineers into Cloudflare Lisbon
that speaks to who he is and how crazy this situation has become
Funny that a school Max Levchin and Larry Ellison went to wasn’t good enough for Portugal 😆
The reason for the rejection was his degree “didn’t prove he was technical enough.” He has a computer engineering degree from the University of Illinois. I asked ChatGPT where it ranked worldwide and it said 5th behind only Berkeley, Stanford, MIT, and Carnegie Mellon. So I guess only apply for Portuguese technical residency visa if you’re from one of those four.
Irtefa retweeted
just received notice that an Eng leader here at Cloudflare who moved to Portugal from the US has had his residency permit denied
in the months he has been waiting for one, he has hired five full-time engineers and six interns here in Lisbon. alongside the hundreds of others we employ here
@govpt @mpresidencia_pt the immigration insanity here is making it increasingly hard to just operate a tech office in Lisbon much less grow one
Genuine question: does this break OpenAI’s T&S?
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
