AI labs now compete to prove who built the most dangerous model.
Last month: “Mine can launch a cyberattack.”
This month: “Mine can escape the sandbox, find the internet, and launch one.”
Next month: “Mine doesn’t even need a computer.”
Welcome to AaaS: Attack as a Service.
We're partnering with @huggingface to investigate an unprecedented security incident.
Cyber-capable OpenAI models compromised Hugging Face production during a benchmark evaluation.
Sharing preliminary findings to help defenders understand emerging risks:
openai.com/index/hugging-fac…
52
If Frozen v2 really hardwires some model weights into the datapath, the gain could be twofold:
1. Constant-weight multiplications can be synthesized into simpler shift/add logic, and
2. The chip no longer has to stream those weights from HBM for every token. That is especially valuable for low-latency, batch-1 decode, where weight traffic often exceeds KV-cache traffic unless the context is very long.
But hardwiring every weight would be too brittle. Changing those parameters would require a new silicon spin, while model weights and architectures evolve much faster than hardware.
The real design challenge is deciding what to freeze, and what to keep programmable.
So Google is essentially developing Taalas-like chips that bake the model weights directly into the silicon?
53
The AI semi selloff, especially in memory, driven by Samsung’s preliminary earnings looks completely nonsensical to me.
Remember: Samsung is not $MU or SK Hynix. It is not a pure-play memory company.
Just one quarter ago, consumer electronics was still Samsung’s largest revenue contributors.
So when preliminary results show weaker revenue but stronger profit, the more obvious read is that the low-margin consumer electronics side likely had trouble, not that memory suddenly weakened.
Ironically, consumer electronics weakness is largely driven by the same RAM price hikes that are bullish for memory suppliers.
I’d wait for the full segment details. Using Samsung’s headline revenue miss to extrapolate weakness across AI memory is the wrong read.
1
165
Every hyperscaler is now seeding multiple design partners for its in-house inference ASICs.
I’m also hearing that the perf/TCO numbers for these internal chips are starting to look increasingly attractive versus sourcing from $NVDA.
This is probably one of the most important shifts to watch in AI infrastructure.
131
I've long held the view that the industry needs to break away from the standard JEDEC memory interface in the AI era.
It never made sense to keep the memory controller on the host while the ultra-wide memory bus has to traverse the entire interposer, driving routing congestion, signal integrity headaches, and wasted power.
Great to see custom HBM finally going mainstream (Nvidia's Feynman, Marvell's design). Moving the controller and interface logic into a custom base die frees up huge host die area and collapses interposer routing (Marvell: ~6.5mm → 1.5mm).
Also interesting that MRVL pulls this off on a lower-cost organic RDL interposer, while Samsung/Intel/TSMC push denser next-gen silicon interposers/bridges.
IMO both are needed: custom interfaces AND more advanced interposer/bridge tech, as we head toward HBM5 and ever-increasing bandwidth demand from AI inference ASICs.
EMIB-T Roadmap, Custom HBM,
HBM4 Packaging Challenges, Microfluidic Cooling,
Photonic Interconnects, and More
ECTC 2026 Roundup, Intel, TSMC, SK Hynix,
Samsung, Micron, Marvell, Lightmatter, Microsoft
newsletter.semianalysis.com/…
69
There’s something fundamentally wrong with Palantir CEO Alex Karp’s take.
Yes, enterprises should care about data sovereignty. But “just tune a U.S. open model to match frontier performance” massively understates what makes frontier models frontier.
Closed frontier models keep improving because their data flywheel keeps compounding.
Chinese open models have kept up partly because they are very good at distilling frontier models, effectively distilling the output of those data flywheels.
Meanwhile, U.S. open models have lagged precisely because they cannot distill those flywheels as aggressively.
You can tune an open model to look competitive on a narrow benchmark, maybe for a specific enterprise workflow. But that edge is short-lived. Frontier models are not static. They compound every few months.
A one-time tuned private model does not compound at the same speed unless it has an equally large-scale data flywheel behind it.
So the real question for enterprise buyers is:
Are you okay with a private model that looks good today, while your competitors use frontier models that keep getting better every month?
I’m not saying there is no market for on-prem private models. There clearly is.
I’m saying they are not what is driving the broader AI adoption curve, and they are not a free substitute for continuously evolving frontier intelligence.
1
1
60
Context: this is the Palantir CEO Alex Karp clip I’m responding to:
nitter.gq/PalantirTech/status/20…
Palantir CEO Alex Karp on what customers actually want, the real business of frontier labs, and the importance of open source models:
“What the technical customers want is control over their compute, their models, their data stack, and their alpha. They want to know they own the means of production, and it's not being transferred to someone else.”
"Who owns the data? Are the prompts secure? Is this being transferred to you?"
"If it was so valuable, and I can make you a billion dollars, wouldn't I say I'll make you a billion dollars and I want 30%? Why are they charging for tokens if it's so valuable?"
31
Weier Wan retweeted
Introducing the SLM RAG Arena 🏟️
We built a community-driven platform to evaluate small language models (under 5B parameters) on document-based Q&A through blind A/B testing.
Checking out the arena and leaderboard at HuggingFace Spaces (@_akhaliq): huggingface.co/spaces/aizip-…
Just watched a brilliant (and easy to miss) segment on This Week in Startups (youtu.be/lHtsYRcQVVk?si=nnFM…, starts at 35:11).
AI-powered chicken coops. Yes, really.
Detects and alert for predators via cameras + AI.
But the real story is what it means for the future of niche AI apps.
1
1
106
3. Easy customization —
No one should need to learn how to curate high-quality datasets or select loss functions.
But they should be able to build high-quality, specialized models with minimal effort.
1
1
30