An investor letter obtained by Newcomer describes a $6B non-exclusive license plus a $1B Nvidia investment at a $12B pre-money valuation. The structure prices access to Poolside’s model-building work while leaving the company intact; neither party had published first-party confirmation.
daily notes
previous notes →Themes
No themes have been published for this stream yet.
Must read
GLM-5.3 shifted scaling from parameters to long-horizon RLLatent.Space / Z.ai
In the August 20 interview, Z.ai CEO Jie Tang says GLM-5.3's gains came solely from RL in long-horizon engineering environments and that parameter count is insufficient context. Z.ai's August 14 launch reports Terminal-Bench 2.1 at 88.2 versus 81.0 for GLM-5.2 and delays weights about two weeks pending safety evaluation; the results remain self-reported.
Signals
NVIDIA reports that AVO completed all 183 levels across 25 public-set environments with a 100.00 RHAE score, using about 12% fewer actions than VISTA. NVIDIA explicitly says the comparison is not a controlled ablation and the result excludes the semi-private and fully private sets.
As we continue to push the frontier of capabilities while improving efficiency, we're dropping API and credit pricing of GPT-5.6 Sol by over 20% for the next 3 months. https://t.co/UoTb3hcB2t
discussion1 selected reply
The temporary reduction is live in the API and applies to ChatGPT Work and Codex credits. OpenAI says Pro, Plus, and Business subscription usage is unchanged, making this a targeted serving-price move rather than a broad subscription reset.
For the past 2+ years, I’ve been working on a from-scratch LLM inference engine, with its own MLIR-based compiler. It’s currently faster than vLLM on most GPUs by about 10-20%, and I expect that number to be higher after more compiler work. In terms of feature set, it’s pretty…
For the past 2+ years, I’ve been working on a from-scratch LLM inference engine, with its own MLIR-based compiler. It’s currently faster than vLLM on most GPUs by about 10-20%, and I expect that number to be higher after more compiler work. In terms of feature set, it’s pretty barebones, only supports Qwen3/3.5, FP8/NVFP4 quants, KV cache quant, and dflash. I’m hoping to release it sometime this year, so I’m making this post to gather some feedback on what features people look for in an inference engine. Please comment below some of the features you always use in vLLM/SGLang/llama.cpp/etc, and I’ll see if they’re doable. Could be anything, like KV offloading, hybrid CPU-GPU inference, etc. Thanks.
discussion1 selected reply
End-to-end runs at high context with no prefix caching and high concurrency
The builder disclosed a from-scratch LLM engine with an MLIR compiler, current Qwen3/3.5 and FP8/NVFP4 support, and a self-reported 10–20% lead on most GPUs. No code or benchmark harness is public yet; the post is useful as a feature roadmap, not a validated performance result.
Sentiment
Harnesses advanced; economics stayed unsettled +0.29
5 posts · 88% confidence
Themes
Buyback expansion met a 5.23% 30-year yield
Doubling planned buybacks did not anchor the long end. Treasury will raise each long-end operation to at least $4B from $2B starting September 9, while the 30-year yield finished August 20 at 5.23%; the purchases support liquidity rather than set rates.
Good morning, Asia. While you were sleeping, our most-read story was about 30-year Treasuries being hit with a fresh wave of selling despite the US government saying it would 'at least double' purchases of securities. https://t.co/tRJTlGCVy8 https://t.co/ZbK9KMEUNb
Must read
No must-reads have been published yet.
Signals
A Citadel client letter obtained by CNBC says the firm shed more than 80% of aggregate risk from the Situational Awareness portfolio through over 100 block trades worth more than $4B. Wellington returned 5.94% in July, but the letter does not isolate the portfolio’s profit contribution.
July payroll gains were significant in only one stateU.S. Bureau of Labor Statistics
Preliminary July state estimates showed a significant payroll gain only in Maryland (+11,700) and a loss only in New Jersey (-25,600), with 48 states and D.C. essentially unchanged. The national jobless rate held near 4.1%; preliminary March benchmark revisions arrive August 28.
ATRenew's Q2 growth outran its Q3 guideATRenew Inc.
Unaudited Q2 revenue rose 32.4% to RMB6.609B, GAAP operating income rose 95.7%, and adjusted operating income rose 70.1%. Q3 guidance of 23.1–25.1% growth marks a deceleration, while management says the outlook remains preliminary.
Sentiment
Long-end pressure met faster risk transfer -0.16
4 posts · 89% confidence
Now available on the API and rolling out across eligible plans for ChatGPT Work and Codex credits. Pro, Plus, and Business subscription usage remains unchanged. https://t.co/DzyrOMdXpW