Jonathan Downing

Building @QuaiNetwork @PelagusWallet @KipperMoney @BlipPay

[Jonathan Downing @jdowning82]

@P3b7_ The bug was introduced in a regression a few days ago and is a range proof cache-key serialization collision. It's incredibly unfortunate and I hope the attacker returns the funds.

[Jonathan Downing @jdowning82]

@Rob1Ham The bug was introduced in a regression a few days ago and is a range proof cache-key serialization collision. It's incredibly unfortunate and I hope the attacker returns the funds. We need to be way more careful when we review PRs to open source projects.

[Jonathan Downing @jdowning82]

@zackbshapiro The bug was introduced in a regression a few days ago and is a range proof cache-key serialization collision. It's incredibly unfortunate and I hope the attacker returns the funds.

[Jonathan Downing @jdowning82]

@OrnnExchange Isn’t selling inference far more profitable? A B200 node can do thousands of tok/s on GLM or Kimi. Payback period would be way shorter

[Jonathan Downing @jdowning82]

@aimalysheva @mostik_ai So it’s basically EAGLE speculative decoding without a verifier?

[Jonathan Downing @jdowning82]

@BNyeggen @usr_bin_roygbiv Right, it would be int8, but that’s still pretty good with Intel’s convrot int8 technique

[Jonathan Downing @jdowning82]

@i2cjak >26 weeks lead time

[Jonathan Downing @jdowning82]

@zekramu @usr_bin_roygbiv Nice, gave you a follow. I was also working on AMX but DDR5 prices are just out of control. So I’m doing something similar on Gaudi 2

[Jonathan Downing @jdowning82]

GLM 5.3 is fp8 natively so around 750 GB, but could be quantized. And yes it's open weights, so any company "releasing" an abliterated version of it should release the weights as well. If the model actually finds a bug in your software, you'd want that information to be stored only on your hardware.

[Jonathan Downing @jdowning82]

@LaylaEleira @RonfortMartin Agreed, this model should be open weights. Gatekeeping a model like this is bad. Especially when you advertise it as a "release"

[Jonathan Downing @jdowning82]

@januarycomputer They didn't even release the weights

[Jonathan Downing @jdowning82]

@abliteration_ai You wrote that you’re releasing the model today but where are the weights?

[Jonathan Downing @jdowning82]

@talhof8 Did you have to get access to special versions of GLM 5.3 and GPT 5.6 to avoid the guardrails

[Jonathan Downing @jdowning82]

@fal Are you planning on releasing open weights?

[Jonathan Downing @jdowning82]

@rehan_shei Are you planning on releasing the weights?

[Jonathan Downing @jdowning82]

@insane_analyst A RISC-V CPU with systolic arrays and an extended instruction set to use them with on-die HBM/SRAM could be cool. Well I guess that’s just tenstorrent then, isn’t it

[alan ⚡💵 @0xalank]

Ravencoin (@Ravencoin $RVN) recently suffered a chain split due to a consensus vulnerability.

And while @QuaiNetwork does merge-mine with Ravencoin, Quai was unaffected by the incident since the auxpow of Quai consensus validates separately from the chains it merge-mines with.

But independent Quai miners do make up close to 40% of RVN's GPU hashrate, so we're pitching in as volunteers:

- Spinning up a fresh RVN explorer for the recovery chain: https://t.co/MIM8f8OoCH (syncing from scratch, will be caught up soon)
- Maintaining a node repo with the security fixes: https://t.co/osY6Sn5iYz
- Scanning for further vulnerabilities with GPT 5.6, Fable, and Kimi K3

We're not the RVN team, just builders in the space who want the project to do well and see a good outcome after the chain split.

Open to connecting with others that are part of the recovery effort.

The PoW community is stronger together 🤝⚒️

[alan ⚡💵 @0xalank]

Open, neutral, unstoppable money for eight billion people.

Scaling Proof of Work is the only way to make surveillance money obsolete before it becomes forced. https://t.co/TLmMBFUnCm