edge (@edgedlt) header
edge (@edgedlt) profile picture

retired acc

Joined December 2023
For the record, you’re full of shit
March 2026 compensation to core developers has been disbursed. For the record: the Foundation and NGD, under my management, have funded core developers, contributors, and community groups consistently since inception - through every market cycle, without a single missed period.
5
7
1
23
3,910
Malice, incompetence, what does it matter now? Your leaders were absent, the ecosystem fragmented. Da pulled money out of projects trying to save Neo and poured it into his new grift. A shame the council never had the spine to stand behind me, I gave them enough chances.
2
9
1,095
Playing around with intervention studies on the Grokking (Power et al 2022) mod-add setup. Confirmed that the grokked algorithm is low-rank in Q/K (effectively rank 1–2). Noise tests sensitivity; ablation tests necessity. When they disagree this much, does it mean the algorithm lives in a very specific direction within that low-rank subspace? Q/K turns out to be almost noise-immune here too. Meanwhile the MLP, despite being mostly irrelevant to the task, still takes ~9% damage. Maybe it’s carrying tiny logit nudges or calibration offsets that help generalization but get scrambled by noise? We must go deeper.
3
2
210
Spawned a Tigerbeetle-inspired demo to watch fault tolerance in action and actually found a bug with it (no implicit vote from leader). Works great now. Under chaotic conditions, the smallest networks halt completely, but you can watch resilience improve as you crank the scale.
1
107
AI is fun but TLA+ is pretty hot stuff too 🫣
1
140
Side quests within side quests
1
116
mfw I learned positional information from tokenized input sequences are encoded using angular frequency in the seminal transformer architecture
1
3
349
We do a little bit of LLM finetuning
1
10
368
We do a little bit of levelling up @cryptodavidw 🫡
1
1
11
778
IMO the pay/refund mechanic in 4337 is really awkward because it means 1 USDC transfer under the hood does 3x ERC20 calls (1x to pull the fee in the entrypoint, 1x for the actual transfer, 1x for the fee refund), and that's expensive.
3
9
445
Average Joe can't pick a sane default. Transactions would fault inexplicably if they used CalledByEntry everywhere, and Global leaves them a sitting duck for phishing attacks or malicious contracts. Joe has to verify what the dApp has passed to the wallet for signing. Ideally it's CalledByEntry, I hesitate to call anything in this industry "safe" but it's more tightly scoped than any EVM network and most if not all alt-VMs too. Three pieces of information protect Joe: 1. CalledByEntry means the interaction is restricted to this contract. It can't touch other token balances, for example. 2. Unless you trust this dApp/URL/interaction with your entire wallet, do not approve any transactions where the signature has scope Global outside of test environments. 3. If the scope is CustomContracts, verify that you trust each contract listed (e.g. verified source or each is deployed by a known address). If you are unsure, verify with the dApp dev.
1
1
26
Imagine paying to give someone else permission to spend money out of your account Sorry ERC-20, time for higher standards
In a few years, new Ethereum users won't even know what an "approval" transaction is.
4
3
13
726
Replying to @gsmachado @gakonst
Why imagine what is attainable? 👀
1
23