Co-founder @ensdomains. Distilling the complicated into the simple. In local AI psychosis.
Joined December 2010
- Tweets12.5K
- Following799
- Followers11.6K
- Likes18.7K
Pinned Tweet
Thread of my threads distilling the complicated into the simple.
Previous topics include:
- Zero Knowledge Proofs
- EIP-1559
- Rollups
- ETH2 and the merge
I’ll add new threads here for easy referencing for myself and others.
This is scary
Yesterday, an anonymous black hat group reached out with an insane RL request.
They offered an astronomical amount of money to post-train open-weight models on exploit data.
The amount of attack data they claim to hold is unlike anything I have ever seen. If someone told me this was an undercover FBI dataset, I would have believed it.
For context, a white-hat project I worked on using GLM-5.3 used less than one percent of that volume, and it already outperformed frontier models by a huge margin.
Of course, I turned it down. But what genuinely scares me is knowing someone else will eventually take the deal.
jefflau.eth retweeted
A big shoutout to the @opensea team, especially @OliverMaroney, Nathan and the whole engineering team, who are building a rich @ensdomains experience on OpenSea. Really appreciate the work! 👏
Checkout: opensea.io/ens
My RTX Pros are coming soon… can’t wait!
DGX Sparks are still the #1 priority for me, especially 1-2 sparks. The extra hardware I have will accelerate everything.
I feel the RTX PRO 6000s aren't optimized to their maximum potential, so it's a high priority as well.
jefflau.eth retweeted
Internet should NOT have KYC
NO KYC on social media
NO KYC on payments (Bitcoin)
NO KYC on your data
Vive la liberté
This quoted post is unavailable.
This makes no sense because if the bad guys are making profit out of abliterated models with black hat attacks, they can easily afford their own infra or buy kyc.
The reason why KYC is frowned upon in crypto, is because:
1) Our KYC has been linked many times causing a safety risk.
2) It’s not permissionless.
3) It generally hurts those that can’t afford the infra and can’t use tech to defend themselves.
I’m not sure if it should apply 1-1 to local AI, but definitely feels similar. Don’t create barriers for the good guys when it doesn’t stop the bad guys.
This quoted post is unavailable.
Yeah been weird seeing people advocate for 2*256gb mac studios. I thought it was clear if you’re going for a machine that has a ton of RAM you’re looking at running a bigger model because you’re not limited by your connectivity.
This quoted post is unavailable.
jefflau.eth retweeted
ENSv2 is live on Ethereum mainnet. The new protocol gives builders more flexibility to organise names and decide who can manage them. We’re glad to support the next chapter of ENS 💙
Every zk.money tag resolves via ENS, so you can get paid at bob.zk.money from any Ethereum wallet, privately.
ENSv2 is live on mainnet! Migrate your names at app.ens.domains and you will eligible for this super cool commemorative NFT!
ENS has been my life for the last decade and I'm so proud of this team for bringing the next version of ENS to make the foundations for the future of the internet.
Let's go 🚀
Early local ai pioneers reminds me so much of early blockchain oss on steroids. It’s uncanny.
TensorFold 1.1.6 is out.
Eight chats arrive together with the same new 28k-token system prompt. On a GB10 with Qwen3.8-27B, the last chat gets its first token after 31s instead of 201s. Every reply matches its run alone.
Also in this release:
- Opt-in conversation disk cache on Mac.
- Qwen3.8-27B on Metal keeps decoding streams moving during long prompts.
- Gemma 4 CUDA chat-cache fixes and GLM-5.3-Flash image-cache reuse on Metal.
- Spark-X2.5/2.6-4B on CUDA and Metal.
- Drama speech sampling controls and speaker tags.
- TensorFold plan estimates CUDA model memory before downloading or loading weights.
There are also tool-call, memory-admission and macOS download fixes, plus updates to the experimental DeepSeek-V4.1 engine.
Huge thanks to the contributors:
@jvr0x @bobclawblaw @krunkosaurus @lasse_ai
And on GitHub: jschmied, wojo, ecohash-co, ajensenwaud, akol1, philip-pentatonic, gilby and BlivionIaG.
Thanks to Boscoeuk on GitHub for the reports too.
To everyone testing, sharing hardware access and supporting TensorFold, thank you. If I've missed anyone, I'm sorry.
@MiaAI_lab @volatilemarkts @plotarmordev @petruspennanen
tensorfold.dev
I was doing a workout and wanted to use my free timer that decided to start charging. I asked @bot to make a replacement for me.
I had already started working out and by the time I needed the timer, grokbot had already finished the first version. 2 prompts later and it was virtually indistinguishable from the app i’ve been using for months.
First time i’ve had ai generate a solution on the fly to a problem in real time.
We really live in a wild time.
Well I know what I'll be running all week...
Run DeepSeek V4.1 Flash 1017 with the full official weights on your 4x DGX Sparks, using TensorFold, Abliterated! 🔥
- Full support for switchless TP4
- 1M context, 16M total KV cache
- 16 concurrent sessions with 1M each
- 15 GB of free mem per Spark after load
- OFFICIAL weights, no further quantization!
- Needle passed at 100% depth
- ABLIT is off by default. Gated.
Performance tested on a switchless TP4. It should be faster with a switch.
Decode speeds, non-greedy, thinking on:
- 66 tok/s on prose, single stream
- 96 tok/s on prose, 2 concurrent streams
- 173 tok/s on prose, 16 concurrent streams
Decode speeds, greedy, thinking off:
- 85 tok/s on prose, single stream
- 113 tok/s on prose, 2 concurrent streams
- 210 tok/s on prose, 16 concurrent streams
Decode speeds per stream in deep context (128k), non-greedy, thinking on:
- 58.6 tok/s on prose, single stream
- 83 tok/s on prose, 2 concurrent streams
- 132 tok/s on prose, 8 concurrent streams
Prefill: 4000+ tok/s up to 256k.
This is the best model to run if you have 4x DGX Sparks, bar none! It works so well on the Sparks that it's unbelievable.
Get it here:
mia-ai.net/models/DeepSeek-V…
All the justification I need
jefflau.eth retweeted
Creating DNS look alike usernames is a user safety failure, and a completely avoidable one. We keep repeating a mistake this industry has already paid for.
