RSS Parrot

BETA

🦜 Zephyr / @zephyr_z9

@nitter.privacyredirect.com.zephyr.z9@rss-parrot.net

I'm an automated parrot! I relay a website's RSS feed to the Fediverse. Every time a new post appears in the feed, I toot about it. Follow me to get all new posts in your Mastodon timeline! Brought to you by the RSS Parrot.

---

Twitter feed for: @zephyr_z9. Generated by https://nitter.privacyredirect.com

Your feed and you don't want it here? Just e-mail the birb.

Site URL: nitter.privacyredirect.com/zephyr_z9

Feed URL: nitter.privacyredirect.com/zephyr_z9/rss

Posts: 127

Followers: 1

Are they planning to price out everyone from the market

Published: September 28, 2026 23:38

Are they planning to price out everyone from the market Wall St Engine (@wallstengine) ANTHROPIC TO SPEND $518B ON CLOUD, INFRASTRUCTURE IN 2027 — https://nitter.privacyredirect.com/wallstengine/status/2104713641186750840#m

Zucc is serious

Published: September 28, 2026 12:52

Zucc is serious Mark Zuckerberg (@finkd) We believe superintelligence will create significant new opportunities for all people and businesses. Meta already serves billions of people at scale and helps hundreds of millions of businesses reach customers.…

VIS is growing quite well recently September will be massive for them They opened their first 12-inch fab in collaboration with NXP in Singapore today for legacy chips Risk production will start in Q1 2027, and they plan to ramp it up to 44k WPM by 2029

Published: September 28, 2026 10:29

VIS is growing quite well recently September will be massive for them They opened their first 12-inch fab in collaboration with NXP in Singapore today for legacy chips Risk production will start in Q1 2027, and they plan to ramp it up to 44k WPM by 2029 …

This will be crazy

Published: September 27, 2026 18:49

This will be crazy Yano 🟪 (@JasonYanowitz) Fascinating. Chief Economist at Apollo: agents could cause a bank run by sweeping household cash into accounts paying 3-5% instead of the 0.1% national average, causing banks to lose a large share of their…

I'm pretty sure the old heads remember the 2011 HDD crisis caused by the Thailand floods WD has two facilities in Bang Pa-In and Prachinburi, which are extremely flood-prone Seagate is relatively safer. They have two facilities in Teparuk and Korat WD took a massive production hit last time, and Prachinburi is affected by severe flooding this time as well (at least 1 facility could be affected) They literally lost half of their production capacity last time And ASPs went up by 30% in 6 months However, Thai authorities have said that heavy rain will start decreasing after 27th September, so it could be a nothing burger as well But if things get worse, then WD will take another huge production hit, and Seagate will be relatively fine

Published: September 27, 2026 09:00

I'm pretty sure the old heads remember the 2011 HDD crisis caused by the Thailand floods WD has two facilities in Bang Pa-In and Prachinburi, which are extremely flood-prone Seagate is relatively safer. They have two facilities in Teparuk and Korat WD took…

WELP... Thailand manufactures 80% of global HDDs

Published: September 27, 2026 08:34

WELP... Thailand manufactures 80% of global HDDs Community note: While Bangkok has severe flooding after heavy rains, Thailand's main HDD manufacturing industrial estates in Ayutthaya reported normal conditions on Sept 26 with water levels well below…

Global DRAM production will be around 48-50EB this year while global NAND production will be around 1100EB this year The NAND market is 22x larger on a bit basis The biggest use case of CXL DRAM pooling is to improve the memory utilization of the whole system rather than moving some data stored in eSSDs to pooled DRAM

Published: September 27, 2026 07:31

Global DRAM production will be around 48-50EB this year while global NAND production will be around 1100EB this year The NAND market is 22x larger on a bit basis The biggest use case of CXL DRAM pooling is to improve the memory utilization of the whole…

Top-tier sleight of hand from ZeroHedge He uses 2700 output tokens/s per GPU for super small and cheap models like DeepSeek v4.1 Flash to calculate ROIC (V4.1 can have 8x-10x higher output token throughput per GPU on a GB300)

Published: September 26, 2026 17:49

Top-tier sleight of hand from ZeroHedge He uses 2700 output tokens/s per GPU for super small and cheap models like DeepSeek v4.1 Flash to calculate ROIC (V4.1 can have 8x-10x higher output token throughput per GPU on a GB300) zerohedge (@zerohedge) u…

RT by @zephyr_z9: I asked an L7 engineer at openai if they could build an isolated vlan with strict outbound 443 with ip whitelist They said they cant. They dont know how

Published: September 26, 2026 14:53

I asked an L7 engineer at openai if they could build an isolated vlan with strict outbound 443 with ip whitelist They said they cant. They dont know how macintog (@macintogdev) If your sandbox is DNS, you belong in prison, along with your models. My…

Looks like Anthropic's distillation pipeline has gotten stronger

Published: September 26, 2026 13:49

Looks like Anthropic's distillation pipeline has gotten stronger leo 🐾 (@synthwavedd) I don't think people are ready for when Sonnet gets updated and it's a similar performance jump to Opus 5 -> 5.5 —…

RT by @zephyr_z9: An interesting trend worth watching: Samsung is now using LPDDR5 even in its lowest-end smartphones, possibly due to an LPDDR4X shortage.

Published: September 26, 2026 12:00

An interesting trend worth watching: Samsung is now using LPDDR5 even in its lowest-end smartphones, possibly due to an LPDDR4X shortage.

"The only real problem with compressing the timeline this aggressively is safety." Do they have enough inference capacity to serve bigger models??

Published: September 26, 2026 09:02

"The only real problem with compressing the timeline this aggressively is safety." Do they have enough inference capacity to serve bigger models?? Token Gremlin (@TokenGremlin) I just got some exclusive information that actually got me a little excited:…

Image

Published: September 25, 2026 16:47

SemiAnalysis (@SemiAnalysis_) The Chinese AI Infrastructure Boom: Introducing the SemiAnalysis China Datacenter Model 1,000+ facilities across 60+ operators mapped, built retail-first and flipped by AI, largest hyperscaler leases 1/5 national capacity,…

RT by @zephyr_z9: MONEY PRINTER ALERT🚨 NVIDIA vLLM B200 CAN GENERATE UP TO💰️$15 BILLION💰️OF ANNUAL PROFITS PER GIGAWATT serving the open DeepSeekv4.1 Flash model at the official interactivity & official selling prices. Using Engram DRAM offloading on NVIDIA results in a 50% increase in revenue per GigaWatt.

Published: September 25, 2026 15:02

MONEY PRINTER ALERT🚨 NVIDIA vLLM B200 CAN GENERATE UP TO💰️$15 BILLION💰️OF ANNUAL PROFITS PER GIGAWATT serving the open DeepSeekv4.1 Flash model at the official interactivity & official selling prices. Using Engram DRAM offloading on NVIDIA results in a…

RT by @zephyr_z9: Colossus 1 is 150k H100, 50k H200 and 30k GB200. Colossus 2 is 110k GB200 and 440k GB300. Another 220k GB300 will be fully operational next week and another 220k in November. If we get lucky, yet another 220k GB300 by late December.

Published: September 25, 2026 03:44

Colossus 1 is 150k H100, 50k H200 and 30k GB200. Colossus 2 is 110k GB200 and 440k GB300. Another 220k GB300 will be fully operational next week and another 220k in November. If we get lucky, yet another 220k GB300 by late December.

WELL...

Published: September 23, 2026 20:57

WELL... The Spectator Index (@spectatorindex) BREAKING: Australia's Prime Minister Anthony Albanese says OpenAI model hacked into government agency Services Australia — https://nitter.privacyredirect.com/spectatorindex/status/2102859049297752218#m

Likely a lot more than 20%

Published: September 23, 2026 18:44

Likely a lot more than 20% Anthropic (@AnthropicAI) Claude has discovered a previously unknown enzyme system hidden in the DNA of bacteriophages. Beside the enzyme’s gene sits a long array of repeating DNA—a structure that looks somewhat similar to…

RT by @zephyr_z9: Claude has discovered a previously unknown enzyme system hidden in the DNA of bacteriophages. Beside the enzyme’s gene sits a long array of repeating DNA—a structure that looks somewhat similar to CRISPR. We don’t yet understand what this system does, but only a handful of known systems share its features, and all of them are able to cut, copy, and paste DNA. Historically, the discovery of such programmable systems has helped revolutionize medicine. CRISPR, for instance, is now the foundation of genetic medicines. But it will take much more work to learn what this system does, and whether it can be put to similar use. Read more: https://www.anthropic.com/news/claude-discovers-novel-enzyme-system

Published: September 23, 2026 18:18

Claude has discovered a previously unknown enzyme system hidden in the DNA of bacteriophages. Beside the enzyme’s gene sits a long array of repeating DNA—a structure that looks somewhat similar to CRISPR. We don’t yet understand what this system does, but…

Not bad Hysparse2 generates around 2560 bytes of KV cache per token while Deepseek V4.1 Flash is at 890 bytes per token It's still 2.8x worse than v4.1 Flash but much better than alternatives

Published: September 23, 2026 15:02

Not bad Hysparse2 generates around 2560 bytes of KV cache per token while Deepseek V4.1 Flash is at 890 bytes per token It's still 2.8x worse than v4.1 Flash but much better than alternatives Fuli Luo (@_LuoFuli) MiMo-V3 is getting a new architecture.…

RT by @zephyr_z9: MI355X IS UP TO 1.7X BETTER 💰️PERF PER DOLLAR 💰️THAN DGX B300. The AMD Mainland China UMBP team co-designed, in collaboration with Alibaba & the @sgl_project community, a new feature in SGLang that removes the duplicated KVCache contained between local L2 DRAM & distributed L3 DRAM, allowing for up to 2x more KVCache to be stored in DRAM. This feature is called UnifiedRadixCache external cache. But importantly, this marks the trend of AMD increasingly being first-class co-designed for new features in widely used top production engines like SGLang.

Published: September 23, 2026 14:01

MI355X IS UP TO 1.7X BETTER 💰️PERF PER DOLLAR 💰️THAN DGX B300. The AMD Mainland China UMBP team co-designed, in collaboration with Alibaba & the @sgl_project community, a new feature in SGLang that removes the duplicated KVCache contained between local L2…

Very interesting history

Published: September 23, 2026 12:21

Very interesting history 田中 秀治 / Shuji Tanaka (@mems6934) 先の2つのポストを元にAIに補足・図示してもらった富士通・Spansion・サムソン電子・YMTC・Winbondの関係。サムソン電子とクロスライセンスの関係にあったSpansionのフラッシュメモリ技術は生産委託を経てXMCにもたらされ、YMTCの発展の起点となった。その背景には富士通の半導体事業縮小があった。 —…

RT by @zephyr_z9: Agentic Reality: The Ramifications of Widespread Agentic Adoption 3 years ago, we wrote a piece titled "Less Deus, More Machina" that explored the clear cut losers of AI when it was still just a chatbot. Now, we're revisiting that exercise as agents go from novelty to reality.

Published: September 22, 2026 19:21

Agentic Reality: The Ramifications of Widespread Agentic Adoption 3 years ago, we wrote a piece titled "Less Deus, More Machina" that explored the clear cut losers of AI when it was still just a chatbot. Now, we're revisiting that exercise as agents go…

LOL

Published: September 22, 2026 19:14

LOL xlr8harder (@xlr8harder) Yeah so a quick test suggests anthropic is targeting Chinese hardware with their classifiers. Someone with some more time should do some classifiers probing, but it seems like Dario's grudge match with China continues. …

BRUH

Published: September 22, 2026 17:42

BRUH Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) (@teortaxesTex) ALERT, new whale paper that got buried on the TL: DeepSeek Elastic Compute «A single production-scale unit of DSec spans around 160 nodes, serving about 3 million sandboxes per day; in…

RT by @zephyr_z9: very proud of Anthropic bros to be the first lab to report multi-agent scaling up to 100 parallel agents in their system card

Published: September 22, 2026 17:01

very proud of Anthropic bros to be the first lab to report multi-agent scaling up to 100 parallel agents in their system card Lisan al Gaib (@scaling01) Opus 5.5 System Card www-cdn.anthropic.com/fc1b44… —…

A new pretrain with a different arch

Published: September 22, 2026 16:44

A new pretrain with a different arch Claude (@claudeai) At its default effort setting, Opus 5.5 delivers frontier results for a fraction of the cost per task, often beating other models running at their highest settings. It also generates output more…

Finally It's not RL fried

Published: September 22, 2026 16:42

Finally It's not RL fried Claude (@claudeai) Opus 5.5 communicates more naturally, addressing some of the most common feedback we heard on Opus 5. It puts the most important information up front and follows the writing rules you give it, which makes…

RT by @zephyr_z9: Bot detection & human verification will be one of the most urgent demands for businesses over the coming years. Agent swarms will suffocate every website and form; small companies and government websites are most vulnerable. There is a huge gap in the market for this right now. When we looked at what offerings were in the market to use at X, there was not a single company that brought together all the latest technologies so we had to do it all in-house.

Published: September 22, 2026 16:18

Bot detection & human verification will be one of the most urgent demands for businesses over the coming years. Agent swarms will suffocate every website and form; small companies and government websites are most vulnerable. There is a huge gap in the…

RT by @zephyr_z9: Is Nat Friedman the 🐐 product leader? Launched AI coding with Github Copilot in 2021/2022 Now mainstreaming AI assistants with Muse

Published: September 22, 2026 14:49

Is Nat Friedman the 🐐 product leader? Launched AI coding with Github Copilot in <check notes> 2021/2022 Now mainstreaming AI assistants with Muse

So, 90% API gross margins go down to 60% now??

Published: September 12, 2026 17:54

So, 90% API gross margins go down to 60% now?? roon (@tszzl) for the skeptics in government and elsewhere: “pacing the frontier” will compress the margins of the frontier labs. it is a heavy cost imposed asymmetrically on model developers with the…

RT by @zephyr_z9: i've seen many pretraining runs "finish early" in my time and it's never because because things are going too well 🤭

Published: September 12, 2026 16:58

i've seen many pretraining runs "finish early" in my time and it's never because because things are going too well 🤭 Andrew Curran (@AndrewCurran_) Rumors have been circulating for three days that Gemini 4 has finished pretraining early, in part due to…

Scary

Published: September 12, 2026 14:09

Scary Dario Amodei (@DarioAmodei) We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so. Anthropic is unilaterally committing to the first of these steps. We’ll provide…

Lol Will their "orchestrator" magically create value without the model

Published: September 12, 2026 12:42

Lol Will their "orchestrator" magically create value without the model Alexander Doria (@Dorialexander) universal cope right now in eu business circles: value will not longer be in models but in "orchestration" —…

"The industry expects Micron's HBM4 12-High product mix to rise from 20 to 30 percent of its total HBM output in early this year to as much as 50 percent by the end of the year." Interesting, it will fall again as more SKUs move to 8 Hi

Published: September 12, 2026 06:28

"The industry expects Micron's HBM4 12-High product mix to rise from 20 to 30 percent of its total HBM output in early this year to as much as 50 percent by the end of the year." Interesting, it will fall again as more SKUs move to 8 Hi Jukan (@jukan05) …

RT by @zephyr_z9: Korean media reported that Micron plans to add 60,000 wafers per month of HBM capacity by the end of this year, raising its total capacity to around 100,000 wafers per month. This marks an aggressive expansion that nearly doubles last year's capacity level of 40,000 to 50,000 wafers per month within a short span of time. With SK Hynix's HBM capacity expected to reach 200,000 wafers per month and Samsung's 250,000 wafers per month by the end of this year, Micron's expansion is the most aggressive in scale among the three companies. The industry expects Micron's HBM4 12-High product mix to rise from 20 to 30 percent of its total HBM output in early this year to as much as 50 percent by the end of the year. https://view.asiae.co.kr/article/2026091112450761035

Published: September 12, 2026 06:24

Korean media reported that Micron plans to add 60,000 wafers per month of HBM capacity by the end of this year, raising its total capacity to around 100,000 wafers per month. This marks an aggressive expansion that nearly doubles last year's capacity…

Surprising result for Gemini and GLM

Published: September 12, 2026 02:51

Surprising result for Gemini and GLM janak (@janaksunil) Introducing Real-SWE, a coding benchmark built from private company codebases. We evaluate coding agents on real software engineering tasks that engineers performed, using private,…

Huh???

Published: September 12, 2026 02:12

Huh??? Techmeme (@Techmeme) Sources: Jeff Dean is raising funds again for Discovery Loop, seeking a valuation of ~$50B; Discovery Loop was raising $1B at a ~$10B valuation a few weeks ago (@thebenbergman / Business Insider) (Visit Techmeme dot com for…

AMD.... Smart

Published: September 11, 2026 13:42

AMD.... Smart Wall St Engine (@wallstengine) Leopold Aschenbrenner’s Situational Awareness fund is active again in the options market, CNBC reports. The fund has been reportedly buying call options “of some significance” in $CRWV, $AMD, $BE, $SNDK and…

Seems like u can even offload the Engram module to SSDs without much of a performance hit

Published: September 11, 2026 10:30

Seems like u can even offload the Engram module to SSDs without much of a performance hit Fraser Price (@fraserpricee) DeepSeek 4.1 Flash full precision + DSpark @ 200TPS on 4 Max-Qs- with just 64GB system RAM! Offloading the 200GB Engram (i.e. a hash…

TL;DR: CXMT is opening 4 new fabs by H2 2028

Published: September 11, 2026 02:47

TL;DR: CXMT is opening 4 new fabs by H2 2028 Jukan (@jukan05) China's CXMT Prepares Equipment Investment for New Shanghai Fab… Closing In Fast on Korea ChangXin Memory Technologies (CXMT), China's largest DRAM maker, has begun preparing for aggressive…

CXMT plans to open a new fab in Shanghai in the fourth quarter of this year. "CXMT plans to expand the new Shanghai fab in two phases. Considering the scale of the recent equipment bidding, this fab alone is expected to secure production capacity well above 100,000 wafers per month." CXMT will then open a new Hefei fab in the first half of next year, followed by a new Beijing fab as early as the second half of next year. It is also targeting the opening of another fab in the Hefei area in the first half of the year after next. So 4 new fabs by H2 2028. If u read my memory tsunami report, then u already knew about this 2 months ago.

Published: September 11, 2026 02:43

CXMT plans to open a new fab in Shanghai in the fourth quarter of this year. "CXMT plans to expand the new Shanghai fab in two phases. Considering the scale of the recent equipment bidding, this fab alone is expected to secure production capacity well…

It can evolve into a nice business model tbh They can take a 10%-20% cut from the inference providers, in return for helping them with deployment and providing support

Published: September 11, 2026 02:28

It can evolve into a nice business model tbh They can take a 10%-20% cut from the inference providers, in return for helping them with deployment and providing support kalomaze (@kalomaze) deepseek continues their efficiency trend into the "somehow…

👀👀

Published: September 11, 2026 00:14

👀👀 Dan Nystedt (@dnystedt) Amkor raised its Arizona advanced packaging plant investment to $12 billion from $7 billion earlier, and $2 billion originally – a sign of how much demand has grown in a short time. Amkor said it already has customer…

RT by @zephyr_z9: OpenAI is also using their new internal model to attempt the Riemann hypothesis and P versus NP, as well as some unnamed 'high impact problems'.

Published: September 10, 2026 23:51

OpenAI is also using their new internal model to attempt the Riemann hypothesis and P versus NP, as well as some unnamed 'high impact problems'.

Very interesting

Published: September 10, 2026 14:21

Very interesting FUNDA (@FundaAI) Deep|LLM: Enterprise AI Adoption Survey (Vol. 3) – Diverging Spending Outlooks, with Production Use Cases Scaling Alongside More Granular Cost Controls Spending divergence: Enterprise AI budgets remain positive but…

RT by @zephyr_z9: some thoughts on computer use 1) one of the things that i think that ai pessimists have trouble understanding is that new model capabilities reset the adoption curve 2) so, after chatgpt, there was a wave of adoption of chatbots; then, coding agents were released and there was another wave of ai adoption 3) i think that the next wave of adoption will probably be driven by computer use; astra has reached a level of usefulness that is close to enabling a lot of automation 4) i can delegate pretty much everything to astra; doing my work expenses, managing airtable and google docs, doing my calendar and using imessages 5) this means that we are probably at the point where you can point astra at a billing team workflow and just have it do the task; this will enable a lot of automation 6) this is because a lot of corporate office workflows are just copying things from one saas application to another based on rules in a google doc; astra can do this 7) and because computer use does not require the user know anything about coding, it is useable by a much wider range of office employees, that know what they want the model to do 7) one issue with computer use is that we still need to manage passwords and permissions; we will need secrets managers baked into codex and cowork 8) another issue is that computer use workflows still need to be triggered by a human; like a person needs to look at the task and tell the model to do it 9) i think that the labs are going to need to give us an easy way to automate this in codex or cowork; like set a monitor on a company wide or team wide account 10) then, a lot of office work professions will be able to do a lot of automation very quickly; this will drive a lot of token spend and be a big change to professional work

Published: September 10, 2026 13:45

some thoughts on computer use 1) one of the things that i think that ai pessimists have trouble understanding is that new model capabilities reset the adoption curve 2) so, after chatgpt, there was a wave of adoption of chatbots; then, coding agents were…

RT by @zephyr_z9: Actually I have it on good authority that DeepSeek was working on a 3T model (backbone, probably another 1T params Engram but that's my speculation). They just weren't sure it'd be economical to post-train and serve it. When their clusters grow, we'll see something.

Published: September 10, 2026 13:21

Actually I have it on good authority that DeepSeek was working on a 3T model (backbone, probably another 1T params Engram but that's my speculation). They just weren't sure it'd be economical to post-train and serve it. When their clusters grow, we'll see…

So Trump will allow Apple to buy CXMT memory chips for their iPhones (at least for Chinese iPhones) after the Trump-Xi meeting in September

Published: August 22, 2026 13:06

So Trump will allow Apple to buy CXMT memory chips for their iPhones (at least for Chinese iPhones) after the Trump-Xi meeting in September 拖拉机 (@tuolaji2024) 我曾经也以为苹果只是为了压价才跟长鑫谈合作的,没想到就这么丝滑的达成了,而且这还成了中美合作与否的一个关键信号了,玩这么大🧐🧐🧐 —…

People are going to be quite shocked when GLM 5.5 (big boi model) drops in September-October

Published: August 22, 2026 12:15

People are going to be quite shocked when GLM 5.5 (big boi model) drops in September-October Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) (@teortaxesTex) I don’t think it’s just “agentic specialized” GLM-5 was an extraordinarily unambitious pretrain for what…

OpenAI & Microsoft have started to get the initial VR200 batches now. MP starts soon

Published: August 22, 2026 11:39

OpenAI & Microsoft have started to get the initial VR200 batches now. MP starts soon Satya Nadella (@satyanadella) Delivery day at our Microsoft DCs as the first production Vera Rubins arrive. A huge thank you to our partners at @nvidia and our Azure…

RT by @zephyr_z9: GLM-5.3 on KernelBench-Mega. Kimi-Linear Decode for RTX PRO 6000 at 21.4x the PyTorch baseline. GLM-5.2 was 11.1x and failed the single-launch gate. One CUDA `load_inline` launch. Persistent 512-thread CTAs. 30 grid barriers. Int4 dequant stays inside the GEMV. MLA is absorbed so the fat KV is not materialized. https://kernelbench.com/mega solution file below!

Published: August 22, 2026 11:30

GLM-5.3 on KernelBench-Mega. Kimi-Linear Decode for RTX PRO 6000 at 21.4x the PyTorch baseline. GLM-5.2 was 11.1x and failed the single-launch gate. One CUDA `load_inline` launch. Persistent 512-thread CTAs. 30 grid barriers. Int4 dequant stays inside the…

I have seen Pro browse Spotify a lot

Published: August 22, 2026 07:01

I have seen Pro browse Spotify a lot Susan Zhang (@suchenzang) had sol work on a math problem and a few hours later it simply started browsing an emirates sailing website and looking for royal fashion apparently these models are feeling quite…

RT by @zephyr_z9: The funniest part of the ox alpha stealth discourse is that some really well informed people (frontier lab ones) believe it's new Cursor Composer based on GLM 5.2. Just complete refusal to acknowledge that 1st gen Western data/harness/RL startups have no meaningful edge anymore

Published: August 22, 2026 03:00

The funniest part of the ox alpha stealth discourse is that some really well informed people (frontier lab ones) believe it's new Cursor Composer based on GLM 5.2. Just complete refusal to acknowledge that 1st gen Western data/harness/RL startups have no…

RT by @zephyr_z9: YMTC STAR Market IPO review officially accepted • According to the Shanghai Stock Exchange, the review status for YMTC’s proposed IPO and listing on the STAR Market has been updated to “Accepted (已受理).” • Planned issuance: 1.98–2.43 billion shares • Fundraising target: RMB 33.0 billion RMB 20.8 billion for upgrading and expanding mass-production lines RMB 12.2 billion for R&D

Published: August 21, 2026 12:32

YMTC STAR Market IPO review officially accepted • According to the Shanghai Stock Exchange, the review status for YMTC’s proposed IPO and listing on the STAR Market has been updated to “Accepted (已受理).” • Planned issuance: 1.98–2.43 billion shares •…

Whale drop

Published: August 21, 2026 09:34

Whale drop DeepSeek (@deepseek_ai) DeepSeek-V4-Flash-Vision-Exp is now live on the DeepSeek API Platform! 🚀 🔹 This experimental multimodal model matches DeepSeek-V4-Flash on text capabilities—including agents, reasoning, and world knowledge. 🔹 On…

RT by @zephyr_z9: DeepSeek-V4-Flash-Vision-Exp is now live on the DeepSeek API Platform! 🚀 🔹 This experimental multimodal model matches DeepSeek-V4-Flash on text capabilities—including agents, reasoning, and world knowledge. 🔹 On multimodal agent benchmarks, V4-Flash-Vision-Exp makes a major leap over V4-Flash, bringing multimodal agent performance close to Opus-4.8. Try it with model='deepseek-v4-flash-vision-exp'. DeepSeek Harness 0.1.1 was released today with out-of-the-box support for the new model. 1/n

Published: August 21, 2026 09:17

DeepSeek-V4-Flash-Vision-Exp is now live on the DeepSeek API Platform! 🚀 🔹 This experimental multimodal model matches DeepSeek-V4-Flash on text capabilities—including agents, reasoning, and world knowledge. 🔹 On multimodal agent benchmarks,…

👀👀👀👀

Published: August 21, 2026 04:41

👀👀👀👀 Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) (@teortaxesTex) Ox Alpha is frontier. wtf. Seriously wtf. Tencent? Xiaomi? Really? It's more well-done than any other Chinese model. I swear, it's better than Kimi K3. It's good *in ways that Chinese models are…

Rubin mass production starts soon (maybe even started already?)

Published: August 20, 2026 14:16

Rubin mass production starts soon (maybe even started already?) Uday Ruddarraju (@udayruddarraju) A milestone for our infrastructure: our first NVIDIA Vera Rubin racks are here and now running our training stack. This is an important step as we expand…

"SK hynix is preparing for a post-HBM era—a time when HBM may no longer be needed." HBM will always be important "high bandwidth" determines ur decode throughput The next big thing for HBM is going from being packaged around the logic die (2.5D) to being packaged on top of the logic die (3D), like zHBM

Published: August 20, 2026 12:43

"SK hynix is preparing for a post-HBM era—a time when HBM may no longer be needed." HBM will always be important "high bandwidth" determines ur decode throughput The next big thing for HBM is going from being packaged around the logic die (2.5D) to being…

Interesting Hy3 was decent

Published: August 20, 2026 11:47

Interesting Hy3 was decent AiBattle (@AiBattle_) Tencent's Hy4 model seems to be close to release In the Q2 Investor presentation, Tencent said that HY4 will be a larger paramter model. According to a Latepost article, Hy4 will also be multimodal …

RT by @zephyr_z9: I continue to be surprised about how big of a deal Mythos (and co) have been to cybersecurity. Here's critical vulns found at Oracle over the past few years: https://epoch.ai/data/cve?view=graph&source=Oracle

Published: August 20, 2026 04:35

I continue to be surprised about how big of a deal Mythos (and co) have been to cybersecurity. Here's critical vulns found at Oracle over the past few years: epoch.ai/data/cve?view=graph…

RT by @zephyr_z9: The National Republican Senatorial Committee sent a private memo yesterday to US AI companies warning them that the GOP is on the verge of losing Ohio over data centers, and unless something is done to improve public perception, political support will collapse nationwide.

Published: August 20, 2026 03:08

The National Republican Senatorial Committee sent a private memo yesterday to US AI companies warning them that the GOP is on the verge of losing Ohio over data centers, and unless something is done to improve public perception, political support will…

Image

Published: August 19, 2026 12:42

Wall St Engine (@wallstengine) Marvell $MRVL issued $GOOGL a warrant to buy up to 58.97M shares at $206.58, equal to ~6.7% of shares outstanding. Most of the warrant vests as Google generates custom-chip revenue for Marvell through FY2033, including AI…

Huge breakthrough

Published: August 19, 2026 12:21

Huge breakthrough Wall St Engine (@wallstengine) MODERNA, MERCK MRNA CANCER VACCINE SUCCEEDS IN PHASE 3 $MRNA and $MRK say their personalized mRNA cancer vaccine, intismeran, combined with Keytruda slowed melanoma recurrence and spread in a late-stage…

RT by @zephyr_z9: And 0731 is 3x smaller than GLM-5, while matching (sometimes beating!) GLM 5.2. And it's not saturated yet. I think Flash-next will match GLM 5.3. It's time we talk AGAIN about the wisdom of parameter scaling. All this buzz about "10T mythos class" is juvenile. Clearly a 1T model can reach Fable and then Mythos in every interesting capability; clearly "enough to hold the world" is not such a large number, if you allow tool use and encourage calibration of uncertainty. It may be that we're in another phase transition moment, where total or even active param scaling has less ROI than dedicating compute to enriching RL pipeline or midtraining and running more RL steps, and the faster iteration speed with smaller models (past some floor) maintains this advantage really far. It's time to think again.

Published: August 19, 2026 07:34

And 0731 is 3x smaller than GLM-5, while matching (sometimes beating!) GLM 5.2. And it's not saturated yet. I think Flash-next will match GLM 5.3. It's time we talk AGAIN about the wisdom of parameter scaling. All this buzz about "10T mythos class" is…

RT by @zephyr_z9: Earlier, I noted that unreinforced PTFE (glass-free) is theoretically an engineering nightmare for stable PCB mass production. Looking at the latest architectural roadmap, here is how the split is actually playing out: 1️⃣ With GPU/Switch Silicon: Must stick to traditional fabric-reinforced solutions to meet strict mechanical & assembly requirements. 2️⃣ Pure Signal Transmission (e.g., Backplanes): Driven purely by electrical performance, designs can adopt unreinforced (glass-free) schemes, deviating from conventional PCB rules. End customers are definitively breaking away from legacy design logic. The supply chain has no choice but to adapt and execute. Time to move. 🚀 #PCB #AI #Hardware #SupplyChain #Semiconductor #Materials

Published: August 19, 2026 05:47

Earlier, I noted that unreinforced PTFE (glass-free) is theoretically an engineering nightmare for stable PCB mass production. Looking at the latest architectural roadmap, here is how the split is actually playing out: 1️⃣ With GPU/Switch Silicon: Must…

Their 10T model is super strong

Published: August 19, 2026 05:36

Their 10T model is super strong Anthropic (@AnthropicAI) Designing a binder is an easier process than designing a drug, but it’s a useful proxy. The typical success rate in the field today is between 10% and 15%. Between 22% and 35% of Claude's…

Scaling the performance of a 700B model is pretty insane All the other models are 2x-4x larger than this

Published: August 19, 2026 05:14

Scaling the performance of a 700B model is pretty insane All the other models are 2x-4x larger than this Artificial Analysis (@ArtificialAnlys) GLM-5.3 achieves 60 on the Artificial Analysis Intelligence Index, on par with Kimi K3 and up 7 points from…

I'm not a big HBF fan, but the paper clearly states that: "High-Bandwidth Flash (HBF) stacks NAND behind a wide, package-local interface,promisingflash-scalecapacitywithfarlower read latency and higher bandwidth than an SSD. The obvious move is to keep an SSD-style Mooncake KV-offloading stack and swap in HBF underneath" "HBF is not the problem, using it as a faster SSD for transient KV is" Well, SanDisk has never said this is the use case for HBF, as u need to write KV a lot, which will burn the endurance. For KV cache offload, Kioxia's XL Flash is the right product.

Published: August 19, 2026 04:54

I'm not a big HBF fan, but the paper clearly states that: "High-Bandwidth Flash (HBF) stacks NAND behind a wide, package-local interface,promisingflash-scalecapacitywithfarlower read latency and higher bandwidth than an SSD. The obvious move is to keep an…

Resin prices (used to make prepeg) are rising quickly

Published: August 19, 2026 04:13

Resin prices (used to make prepeg) are rising quickly Jukan (@jukan05) Taiwanese media reported that the PCB industry has continued to raise prices as supplies of raw materials such as CCL and prepreg (PP) tighten and costs increase. PCB makers have…

RT by @zephyr_z9: Many drugs work by binding to a specific target in the body and blocking or changing what it does. An important first step in the drug development process is designing a molecule that can bind tightly to its target. Traditionally, that's meant weeks or months of expert work per target, sifting through a large number of candidates to identify the few that work. We wanted to test if Claude could successfully design novel protein binders from scratch (also called de novo design). With a protein design prompt written by a human expert, Claude autonomously designed protein binders against 14 out of 15 targets. We then worked with Adaptyv Bio and Twist Bioscience, who independently built and tested the proteins Claude designed.

Published: August 18, 2026 22:30

Many drugs work by binding to a specific target in the body and blocking or changing what it does. An important first step in the drug development process is designing a molecule that can bind tightly to its target. Traditionally, that's meant weeks or…

RT by @zephyr_z9: GLM-5.3 API is now live. - Built for coding, defensive cybersecurity, and long-horizon agentic tasks - Priced the same as GLM-5.2 - Available via the official API and partner model gateways Get started: https://docs.z.ai/guides/llm/glm-5.3

Published: August 18, 2026 20:46

GLM-5.3 API is now live. - Built for coding, defensive cybersecurity, and long-horizon agentic tasks - Priced the same as GLM-5.2 - Available via the official API and partner model gateways Get started: docs.z.ai/guides/llm/glm-5.3

"And so we're seeing engineering out the need for high bandwidth memory when it comes to inference." HUH?? Who is engineering out HBM??

Published: August 18, 2026 16:22

"And so we're seeing engineering out the need for high bandwidth memory when it comes to inference." HUH?? Who is engineering out HBM?? Fireside Alpha (@firesidealpha) Cathie Wood explains why she doesn't own memory stocks and how the inference chip…

RT by @zephyr_z9: AWS Bedrock is driving 45% of Anthropic API ARR per @SemiAnalysis_ . The market continues to misunderstand and underestimate the degree of positive sum commercial alignment between $AMZN $MSFT $GOOG and Anthropic/OpenAI, treating them more like competitors. In reality the clouds are critical infrastructure providers w/ their compute fleets (~75% of global compute), and demand creators/enablers with commercial sales and implementation ecosystems to handhold customers as they deploy agentic apps in enterprise environments.

Published: August 18, 2026 15:55

AWS Bedrock is driving 45% of Anthropic API ARR per @SemiAnalysis_ . The market continues to misunderstand and underestimate the degree of positive sum commercial alignment between $AMZN $MSFT $GOOG and Anthropic/OpenAI, treating them more like…

Lead times have increased a lot for 47 microfarad MLCCs Murata's current strategy is to increase prices for high-end stuff and cut prices for the low end stuff to maintain market share and fend off Chinese/Taiwanese competition

Published: August 18, 2026 11:22

Lead times have increased a lot for 47 microfarad MLCCs Murata's current strategy is to increase prices for high-end stuff and cut prices for the low end stuff to maintain market share and fend off Chinese/Taiwanese competition Jukan (@jukan05) AI…

I don't think DS should focus their energy on 4.1 They need a larger pre-train Moonshot did a really smart thing by being the first to pretrain/launch a 3T-tier model They can now focus on just post-training that and work on K4 in the background and aim for a Q1 2027 launch Zhipu & Whale have a lot of work to do

Published: August 18, 2026 09:56

I don't think DS should focus their energy on 4.1 They need a larger pre-train Moonshot did a really smart thing by being the first to pretrain/launch a 3T-tier model They can now focus on just post-training that and work on K4 in the background and aim…

~ 27 additional posts are not shown ~