Toshii
@0xToshii
Followers
1K
Following
1K
Media
44
Statuses
520
released the (n)th challenge (20+) for my smart contract CTF: Mr Steal Yo Crypto all challenges are based on real world exploits and are well suited for devs exploring #solidity/#security primer & hints: https://t.co/dSjd8j7wkP challenges:
mrstealyocrypto.xyz
Solidity CTF challenges inspired by real world exploits
4
10
49
had a friend at a tech company tell me llms are making ppl hella lazy - a bunch of internal search / recommender solutions being pumped out as basic ootb text rag + llm reranking + llm as a judge testing w synthetic data era of workslop is upon us
0
0
2
Even if LLMs don't improve at all from today, they are going to radically change everything. we are not even close to full adoption
Even if LLMs don't lead to AGI, LLMs will lead to amazing technological breakthroughs, especially since we started to train them more and more with RL "learning by doing".
0
0
3
literally just semantics yet somehow manifests irl
very bullish on ML very bearish on AI
0
0
2
This is going to be so good in the Cline cli Workflow I'm imagining: > cline spawns, implements feature and creates a pr > cline spawns, gathers context from initial run, reviews pr and makes any changes required / adds suggestions > I get a notification and review the final pr
0
1
2
incentivizing insider trading is possibly the greatest modern innovation for gathering knowledge
@ChetasOG There should be a polymarket for betting on which lab made this model
0
0
4
Limited work on coding agent testing w/ synthetic user turns and rl using coding agent scaffolding that incorporates user feedback in os literature This seems like it could be a solid research direction?
Interestingly itβs still a βoneshotβ benchmark, where there are no user turns other than the initial prompt. Less than 1% of all real world coding tasks take on this shape. So while this is a good start, weβre still not close to evaluating / simulating real world AI performance.
0
0
1
excited for my ai drone paramilitary group
2
0
2
History behind this: sonnet4 was overfit to the --/++ diff edit delimiters and was incapable of reliably using the older >>/<< markers, so we swapped to this default format, although our rails handles both code data often uses one of these but would be an interesting exp to run
1
0
3
as an aside the sec review we got was so mid, interesting being on that side
0
0
0
defi project i wrote the smart contracts for a while ago shut down recently wasn't involved in operations. was interesting watching them seemingly barely not getting it to work the way they wanted to, always felt like it could pop off
1
0
0
this would be so hype, honestly having a really good os coding model that can do image -> ui workflows will be such a game changer
Hi, @huybery ,when are you planning to open multimodal Qwen3-Coder?
0
0
2
new idea for my psyop campaigns
such a turnoff when companies do this not because it's unproductive for the TL (or whatever medium you're on), but because it signals bad taste, which extends to product (that said I like firecrawl so this surprises me)
1
0
2
ATTENTION NEW YORKERS π¨ Join us for the world's fastest hackathon, with a 2 hour timeline only possible when powered by @CerebrasSystems inference x Cline. Even better, run the chance to win ππππͺππͺπ¨ (yes!), πΏπͺπππ πππ€ππ€π‘ππ©π, and $1000 in cash prizes! πΉπ«β¨
0
0
3
composable system prompts optimized for major llms is going to go so hard
0
2
10
gm we added a toggle to @cline allowing you to switch between the old smart-truncation context management and our new auto-compact weβve also observed some non-frontier models are kinda getting lost in the sauce with auto-compact, so weβve gated this feature by model (for now)
2
2
48
whoβs making money from ai coding? - bring your own inference app layer (cline, etc) - cost optimization app layer (cursor, etc) - pure play inference (cerebras, etc) - big model labs (anthropic, etc) - inference routers (openrouter, etc) - hardware (nvidia, etc) so on so on
1
4
20