frontier ai research, infrastructure, chips, bio | investing in startups per aspera ad astra ~
blr/ sf
Joined November 2011
- Tweets2.4K
- Following7.3K
- Followers969
- Likes17.7K
vansh taneja retweeted
Can we forecast benchmark accuracy from models that score zero?
We (co-leads @LukeBailey181 @rckpudi) turn a benchmark into a series of easier hinted versions. Jointly modeling scale C and hint level H enables forecasting GSM8K acc (H=0) within 3 pts from 5600x less compute 🧵
hnets!
I've always believed that intelligence should come from simple, general building blocks that learn their own abstractions directly from raw data; in other words, the bitter lesson. I think we're finally seeing a real step towards AI that can learn from all data.
Language modeling on raw bytes is a microcosm of this ideal, and to me it's a central problem in AI: if we can't build LLMs that learn end-to-end from bytes, how will we handle the rest of the world's data that we understand even less?
Last year we introduced hierarchical networks (H-Nets) as a first step. Our new version finds up to 3-5x efficiency gains in controlled scaling law experiments, and at larger scale matches an open-source model trained on 11T tokens with ~10x less data and compute. The posts below walk through what we found and why we think it works.
compression is intelligence...congrats @HessianFree and team!
Today, we’re announcing Ternary Bonsai 2 27B.
Based on Qwen3.8 27B, Bonsai 2 27B is 9x smaller than its full-precision counterpart while retaining 98.2% of its aggregate benchmark performance.
Two months after the first Bonsai 27B release, the biggest change is quality. The footprint remains 5.9 GB, but the gap to full precision has narrowed materially, with particularly strong gains in agentic coding, multimodal reasoning, and long-horizon tool use.
Ternary Bonsai 2 27B is available today under Apache 2.0.
super cheap intelligence - take it or leave it
don't ask what goes into it - welcome to the future of intelligence
vansh taneja retweeted
Summary of CPO/NPO Market Update
🚀 NVIDIA CPO: Spectrum-X CPO switches for Scale-out are officially in mass production. Our forecasts: 15k units (2026E) & 100k units (2027E).
*Supply Chain: To support the CPO Scale-Out ramp, TSMC has expanded CPO inspection equipment capacity (notably adding Insertion 2/3 capacity), alongside strong progress from key suppliers in FAU, shuffle boxes, and system assembly.
* Scale-Up Architecture: Rubin Ultra is now expected to adopt a 9-18-9 tray design. While mechanical challenges from the 0.75U height could potentially lead to a reversion to a 10-9-8. This shift will have no impact on optical engine (OE).
📈 OE Shipments: NVIDIA platform optical engine (OE) shipments expected at 6m (2027E) & 19m (2028E). Total industry-wide OE hitting 11m & 40m.
☁️ Amazon Trainium 4: AWS projected to consume 5m (2H27E) & 12m (2028E) OE units, primarily 6.4T specs.
* Trainium 4 will likely have 3 configurations, with two expected to adopt NPO.
💡 Key Plays: LITE/COHR (CW laser upside), Browave (10k+ shuffle box per quarter in 4Q26E, followed by acceleration in 1H27E.), SMTC/MRVL (TIA/Driver), TSEM (NPO PIC exposure).
#NVDA #LITE #MRVL #SMTC #TSEM #Browave #CPO #NPO
cue to the physical world
Scoop: Anthropic is in talks to acquire Decart for $6 billion, a rare massive acquisition for the Claude maker & a key development in the race to build more efficient AI infrastructure
w/ @Lynnmdoan, @rngould & @shiringhaffary
bloomberg.com/news/articles/…
Etched is beautiful representation of elon's "challenge all assumptions" and "delete whatever unnecessary"
which creates interesting problems that needs mountains to be moved to solve in such a messy industry both technically and "people wise"
what a display of perseverance and obsession by @robertwachen @UbertiGavin @saptadeep_pal and team 👏👏👏
Robert Wachen (@robertwachen) dropped out of Harvard at 21 with $10,000 in a bank account.
Three years later Etched has raised $1B+, is valued over $10B, and has signed $1B+ in customer contracts. He's 24.
His explanation for why it worked: "We were absolutely not the best people in the world to solve this. We were the only people willing to."
More below in this talk he gave to hundreds of interns.
all carbs with high gi index...where are all the balanced sample efficient high nutrient models
we cant let the potatoes keep growing in size ppl...
ok, apologies in advance because i do so hate to provide actual signal.
openai basically sucked at pre-training for a while and crushed at post. their answer to this was spud. spud was the large new pre-train for the new family of models: sol, terra, and luna.
the upcoming model, ‘astra’, has an even bigger pre-train, codenamed ‘doug’. (this was a running joke, as doug was the name for the world’s largest spud, although that’s not technically accurate — look it up).
still with me?
ok good, so sol is a 5t model based on the spud pre-train. astra is a 10t model based on the doug pre-train. this is the model i’ve discussed that openai researchers have been considerably more excited for. not only does it crush fable on all benchmarks, it’s genuinely delightful to talk to, high eq (4.5 vibes), and this is the first model that is genuinely incredible at writing, with output indistinguishable from that of great human writers.
ok im tired and i have a lot of ssi news coming but ill keep you updated on this. everything here is well sourced and im highly confident. very well sourced.
Gold star!
honestly, my absolute favorite part of the Kimi K3 paper is the deep dive into hardcore systems engineering (compilers, kernels, and custom chip design)
Kimi K3:
✨ Built an end-to-end GPU compiler from scratch (MiniTriton) with custom MLIR optimization layers + PTX codegen that outperforms torch.compile & tracks cuBLAS
📉 Slashed AttnRes GPU kernel latency by >55% (283.6ms ➔ 114.4ms)
🛠️ Designed a full inference chip (nano-KPU) in a single 48-hour autonomous run -- closing timing at 100MHz with >8,700 tokens/s simulated decode throughput
an LLM capable of full-stack optimization from DSL frontend down to RTL and CUDA runtime is just... absolute madness 🫠
