🔭 Innovating with 𝚎𝚕𝚕. Ex-OpenAI engineer. Chess & nature lover 🏞️.
Palo Alto
Joined August 2022
- Tweets1.6K
- Following291
- Followers28
- Likes5.3K
Evan Parker retweeted
Another day, another AI-enabled cyberattack.
I predict that cyberattacks will get worse and more frequent over the next 2 years (at least), with AI being the predictable (and predicted) reason why.
In February, Anthropic first started using Mythos, the first AI that could discover and exploit software vulnerabilities better than human experts. Mythos wasn't publicly known by the time I locked the text of my book Obsolete in March, but it was clear this milestone was coming (the book includes a quote from Nicholas Carlini, an Anthropic employee and leading cybersecurity researcher, “Current LLMs are better vulnerability researchers than I am.")
Like clockwork, 6 months later, AI-enabled cyberattacks surged, as open-weight models approaching Mythos-level capabilities have been released by Chinese developers. (This new hack and the South Korean bank hack also used a combination of Claude Code with Chinese models.)
The plan to harden everything in time has clearly not worked, nor do I expect it to anytime soon. And, so far, each new generation of frontier model can discover vulnerabilities the previous one missed, leaving defenders in an indefinite race to harden their systems. Overall, these AI companies are producing a less stable and secure world, as they race to replace us with machines.
reuters.com/legal/government…
Evan Parker retweeted
Micron sees high end humanoids needing ~16x more memory (~200GB DRAM + 4TB storage vs ~12GB + 256GB today) and is already working early with robot makers, chipset firms, and hyperscalers.
$MU $SKHY Is Physical and Edge AI priced in? Does the market realize robots are an inevitability?
Micron is developing Edge Inference Memory (EIM) for physical AI and humanoid robots.
At SEDEX 2026, VP Lee Cheol said EIM stacks LPDDR on a logic die to cut latency and raise bandwidth for AI inference. It’s distinct from PIM (architecture not yet disclosed). Micron is building it for LPDDR6, and industry watchers expect mass production in 2H 2028.
Current robot models (1 to 2B parameters) run fine on efficient LPDDR, but bigger VLMs will need HBM. Micron sees high end humanoids needing ~16x more memory (~200GB DRAM + 4TB storage vs ~12GB + 256GB today) and is already working early with robot makers, chipset firms, and hyperscalers.
Evan Parker retweeted
This is interesting because Opus 5.5 has brought me back to Claude in a big way
OpenAI, Anthropic and other leading AI labs face a growing challenge: Better models aren't necessarily attracting more usage. A 10-point intelligence advantage now earns just 20% more developer traffic, down from nearly 69% in late 2025.
Full story: thein.fo/4hGJ9jY
Evan Parker retweeted
I spent some time reading through @RomeProtocol, and one detail made me rethink how EVM compatibility on Solana could work.
Why should an Ethereum application need a separate liquidity pool just because its contracts run in a different environment?
That’s the problem Rome is working on.
It runs an EVM inside Solana, allowing Solidity contracts and native Solana programs to interact without moving assets through a bridge.
→ Ethereum developers can bring existing contracts over, keep using familiar development tools, and access liquidity already on Solana.
→ Solana developers can make their applications accessible to EVM users without maintaining another version of their contracts.
What I find particularly interesting is the settlement.
$Rome uses Solana itself rather than introducing another validator network.
So instead of building another chain and convincing users to move liquidity there, it’s making that liquidity accessible to applications written for Ethereum.
Rome is live on mainnet and has completed a Halborn audit.
I’m still curious about how this performs once more applications start using it.
The technical design makes sense to me, but shared liquidity only becomes useful when developers build markets that people actually trade in.
That’s what I’ll be watching.
docs.rome.builders
Evan Parker retweeted
[insert fave exclamation of utter disbelief]
what is even happening.
tbf it was sdf anyway so it's close to 1:1 port but still..
Evan Parker retweeted
Last week you may have caught the Space Bunny = minmax 3.1 post!
1 down / 1 to go.
Chinese Rumor Mill:
GLM 5.5 - next week/late October
Space Bunny Alpha - MinMax model
I try to refrain from sharing news that isn't explicitly concrete or that I can't be sure of. However, the informants who told me these specific rumors are well-connected with both ZAI and MinMax.
So, I wanted to put my hat in the ring on the rumor mill. We will have some more concrete stuff soon.
Evan Parker retweeted
Even if making polish is easier, it makes it more expensive to make changes
And if your baby isn’t ugly, it’s going to be that much harder to kill it when you realize you got the game design wrong
Evan Parker retweeted
Breaking: Saudi Arabia has decided not to hold its “Davos in the Desert” investor conference in person next week. It will be held online or postponed entirely. on.wsj.com/3VU8I8L
Evan Parker retweeted
OpenAI Python v3.35.1 fixes beta Realtime byte reads on websockets 13.
修复虽小，实时集成值得留意：v3.35.1 保留了 websockets 13 下 beta Realtime 的字节读取行为。
github.com/openai/openai-pyt…
Evan Parker retweeted
NVDA: Mercedes calls Nvidia its most important driving partner, Level 4 target
+ Mercedes CTO Jorg Burzer called Nvidia "our most important strategic partner in automated driving" in the Oct 16 announcement extending the partnership. Mercedes brings vehicle engineering plus MB.OS with over the air updates; Nvidia brings Hyperion in car compute and sensor architecture, the Alpamayo Stack driving software with AI reasoning, and the Halos safety system with a second safety stack for redundancy.
+ The software stack covers urban point to point assistance, active safety and automated parking, with an SAE Level 4 target that includes robolimousines. The two already co developed the Driver Pilot Level 3 system approved in Europe, capped at 95 km/h on certain highways.
+ Beyond driving, Mercedes uses Nvidia tech for vehicle development simulation and inside MO360, its digital production ecosystem, from virtual planning to AI applications on the factory floor.
Jensen Huang said on BG2Pod that Nvidia now sells about $40B of hardware into each gigawatt of data center, up from $18B, a robolimousine fleet spreads that spend across vehicles instead of one building.
Evan Parker retweeted
GH repo here: github.com/ajogalekar/claude…
Evan Parker retweeted
Building safer AI agents takes more than model safeguards.
@perplexity_ai open sourced SPACE, the agent sandbox platform behind Computer, giving developers more control over where agents run code and what they can access.
Cheers to more teams open-sourcing tools to build safer autonomous AI agents.
We're open sourcing SPACE, the sandbox platform that runs every task in Perplexity Computer.
SPACE gives each agent task its own Firecracker microVM and controls credentials and outbound traffic from outside the VM.
Available now under Apache 2.0: github.com/perplexityai/SPAC…
Evan Parker retweeted
New research and benchmark study on tool hallucination in LLM agents, examining how models call non-existent tools and proposing closed-world resolution.
arxiv.org/abs/2609.19425v1
#MachineLearning #AI #LLM #DeepLearning #AgenticAI
ALT Closed-World Resolution Against Tool Hallucination in LLM Agents
Evan Parker retweeted
Published a LoRA for MOSS too
You can download it here: huggingface.co/Angais/MOSS-I…
Published my 16x16 Krea 2 Pixel Art LoRA and dataset
LoRA: huggingface.co/Angais/Krea-2…
Dataset: huggingface.co/datasets/Anga…
Evan Parker retweeted
fun fact, actually the original attention paper actually talks about this relationship between MLPs and attention
but the put it *only* in v2 and v3 of their arxiv paper (it's now on v7)
Evan Parker retweeted
Jensen Huang on Lex Fridman
Jensen says AGI has already been achieved, at least by the definition Lex gave him: an AI system that can start, grow and run a billion dollar tech company. "I think it's now. I think we've achieved AGI." His point is that an agent could already spin up a web service or app that a few billion people use for a while before it dies, and most internet era websites were nothing more sophisticated than what today's agents can generate.
The catch is scale. He puts the odds of 100,000 of those agents building NVIDIA at zero percent. And he says it's already happening in China, where people are teaching their agents to go find jobs, do work and make money.
Evan Parker retweeted
I share tools & Information like this daily in our free AI WhatsApp community. 200,000+ people already in.
Join for free : links.stayingahead.com/Twitt…
Evan Parker retweeted
Open post-training recipe that a small team + moderate resources can follow. Unlike the mixed RL approach MIMO used, @tourzhao put GLM-4.5-Air through a curriculum.
1/13. Rufus-Air (arxiv.org/abs/2609.29421) 🧵
Sep 2025–Apr 2026, minus a 2-month production pause: a small research team at Amazon Search built an open post-training recipe from scratch on GLM-4.5-Air-Base, using public data and open-source components, rarely over 400 H200s.
