
https://teapawt.org/t.co/czU1vnSSM8 https://teapawt.org/t.co/iI4T7246x0 https://teapawt.org/t.co/VKwPIfZpfO https://teapawt.org/t.co/2Ke8acopLv https://teapawt.org/t.co/Qq2nNf5mmX https://teapawt.org/t.co/h4sxAvwhTs https://teapawt.org/t.co/rv1c7xDGaD https://teapawt.org/t.co/OH4EV4ITfn https://teapawt.org/t.co/N5ZzHTQA4G https://teapawt.org/t.co/k8NVce7fHW
España
Joined May 2018
- Tweets241
- Following0
- Followers75
- Likes12
Studio2C retweeted
🚀 New Scrivibe update is live!
👤 New guest onboarding: type a title, get a full chapter.
📚 Earn free tokens by sharing your bookshelf.
✨ We're now on X, follow @scrivibex
Removing every barrier between an idea and a published book.
Studio2C retweeted
Analysis/Summary of leaked DeepSeek Liang Wenfeng investor Q&A - 4 hours, 42pp, May 2026
chatgpt.com/s/t_6a629e40ffdc…
Continuous Learning and RSI:
The most consequential claim is continuous learning. Liang is effectively saying that present agents are powerful but amnesiac contractors; the next frontier is an AI employee that can spend months learning an organization and retain what it learns. If that capability works, he expects it to accelerate AI research itself. But the transcript offers no technical breakthrough—only a strongly held research bet.
US-China gap; Compute resources:
Liang’s central thesis is that China’s AI gap with the United States is overwhelmingly a compute-access gap, not a talent or algorithmic gap. DeepSeek can remain competitive by operating one or two model generations behind, using radically less compute and emphasizing efficiency. But he is explicit that, with Chinese firms confined to roughly 40–50 billion activated parameters while US labs may be experimenting around 800 billion activated parameters, complete parity is impossible without much more hardware.
His technical roadmap is:
base language model → chain of thought → agents → continuous learning → AI-assisted self-improvement → embodiment
Huawei 950 versus NVIDIA
Liang’s comparison is nuanced, despite the provocative line that NVIDIA is “digging its own grave.”
He says:
four Huawei 950s are approximately equivalent to one NVIDIA GB300; Huawei is about two years behind at the chip level; a Huawei system may cost 50–200% more and still be economically acceptable; Huawei’s useful depreciation life is closer to three years, versus roughly five years for NVIDIA; after a few years, the Huawei parts may become unattractive because of electricity consumption.
Thus “parity” does not mean per-chip parity or energy-efficiency parity. It means that a sufficiently large Huawei system can execute the same workload at tolerable latency and cost.
The transcript’s claim that the Huawei 950 SuperPoD can replace GB200/GB300 is also a system-level claim. Huawei’s full Atlas 950 SuperPoD comprises up to 8,192 Ascend 950DT NPUs in 160 cabinets, with a claimed 8 EFLOPS of FP8 performance. An NVIDIA GB300 NVL72 is a single rack containing 72 Blackwell Ultra GPUs. These are radically different physical scales, so aggregate peak performance does not establish equivalent per-chip performance, total cost of ownership or training efficiency. Huawei’s specifications and NVIDIA’s GB300 specifications confirm the system configurations but do not independently validate Liang’s workload-equivalence claim.
Huawei’s real bottleneck: production
DeepSeek had apparently been allocated 16,000 Huawei 950 chips, which Liang treats as only about 4,000 NVIDIA B-series equivalents. That is enough to reproduce or support the present model generation and develop the software ecosystem, but not enough for DeepSeek’s next frontier model.
His timetable is:
ecosystem usability: potentially demonstrable within about one year; near-term hardware constraint: production capacity, packaging and availability;
2026–28: continued capacity bottleneck;
five-year horizon: he is optimistic that production will cease to be the binding constraint.
So his position is not that Huawei has already erased NVIDIA’s advantage. It is that CUDA lock-in may be eroded much faster than the semiconductor manufacturing gap, leaving capacity and energy efficiency—not software compatibility—as the enduring constraints.
Studio2C retweeted
🚀 New Scrivibe update is live!
👤 Edit any chapter directly in the app
📖 Share your finished book, cover, and trailer in one click and earn reward tokens for doing it
✨ Faster page and content loading
Every update closer to the fastest, smoothest eBook creation tool.
Studio2C retweeted
if you're a startup founder just read this so I can sleep peacefully tonight
most companies can get to $1M ARR with only two channels
the channels you should pick from are
- paid ads
- cold email
- cold DMs
- founder led social
the channels you should not pick from are
- SEO
- email newsletter
- guesting on podcasts
- youtube channel
why
you need revenue tomorrow, list one makes that happen
list two makes revenue happen in the future
they are long term investments that pay off 12 months from now
this is the most common mistake I see founders make
the framework to think about this with is how you think about personal finance
when you're young, 80% of resources go to surviving and 20% to investment
when you're old, 80% goes to investment and 20% to surviving
apply this same idea to your young company
gl hf
"hey Claude code make a plan to do list 1 based on cody schneiders tweets"
Studio2C retweeted
We’ve received notice that the Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5.
We'll begin restoring access tomorrow, and will share an update soon.
We’re grateful to our users for their patience, and to everyone who worked with us on redeploying the models.
Studio2C retweeted
I built a Chrome extension that turns your "read later" list into dedicated reading time on your calendar.
Save 5 articles → it auto-books a 30-min "reading block" on your Google calendar, links included. So you'll actually sit down and go through them.
No account, no server, everything local. Open source (link below)











