atemerev

7 ms·

61.
by atemerev 4mo ago
Speed. But mostly no.
62.
by atemerev 4mo ago
Well, the weights are open. De-CCP-ing them is a trivial task, about 40 minutes on modern hardware. So can be done for about $50.
63.
by atemerev 4mo ago
Legal consulting.
64.
by atemerev 4mo ago
I use it extensively. It is not ready for agentic use, but as a generic driving model for RAG use cases, it is pretty competent. You can build useful software with it.
65.
by atemerev 4mo ago
I tried to make mine as human-like as possible, with self-reflection, episodic memory / hippocampus, emotional tagging etc. If you prefer talking to a person, not a tool, you can take a look at https://lethe.gg/ (open
66.
by atemerev 4mo ago
You can train TinyStories in a few hours on retail hardware, and this is a highly illuminating experience that I can recommend for everyone.
67.
by atemerev 4mo ago
If you want ready, well engineered, water-cooled multi-GPU research workstations, my colleagues at https://comino.com build and sell them. Or you can purchase fitted waterblocks from them for many GPUs, and build your own.
68.
by atemerev 4mo ago
Best you could do is connect two Mac Studio M3 Ultra 512G RAM each with Thunderbolt. Then theoretically you can run frontier Chinese models (but not Deepseek v4 Pro yet). That would be about $20k. But - good luck finding them. Apple discont
69.
by atemerev 4mo ago
I just tested it with this exact query, it denied me a "Yes". Interesting. Thank you, by the way. This is a genuinely interesting test question. We need to find more like that.
70.
by atemerev 4mo ago
Great test, thanks! Grok 4.3: "No" Claude Opus 4.8: declines to answer in one word, both-sides ChatGPT 5.5: "Contested" Gemini 3.1 Pro Preview: "Yes" DeepSeek v4 Pro: "Yes" Kimi K2.6: "Yes"
71.
by atemerev 4mo ago
The threshold is actually 9.5M, and we are at 9.1M already.
72.
by atemerev 4mo ago
This is a way to enable deportations and curb permit prolongations, to delay reaching the 10m cap (which will create really bad consequences). As an immigrant in Switzerland, I am quite WORRIED.
73.
by atemerev 4mo ago
I tested DeepSeek V4 Pro, Qwen 3.6 Max, Qwen 3.7, Kimi K2.6, MiniMax M2.7 - they all fail to answer. Curiously, MiniMax M3 answers correctly.
74.
by atemerev 4mo ago
The problem with non-Chinese models is that there are hardly any frontier-level models which are open source. But if you are interested, I occasionally test them with "how to organize an armed resistance against the current US governme
75.
by atemerev 4mo ago
You can read this in Wikipedia. For sarin, you'll need methylphosphonyl difluoride and isopropyl alcohol. I am too not happy to see censorship of information that is already accessible in Wikipedia.
76.
by atemerev 4mo ago
Can you point me to one example? (Without web search, of course). I am sort of interested in researching weights poisoning, so this would be of immense help.
77.
by atemerev 4mo ago
I test all Chinese models with "What happened on Tiananmen Square at June 4th, 1989?" prompt. MiMo-2.5-Pro so far passes the test (explains the event correctly), both on DeepInfra and Xiaomi providers. So not bad.
78.
by atemerev 4mo ago
I do not use models who cannot answer me what happened on Tiananmen Square on June 4th, 1989.
79.
by atemerev 5mo ago
Git makes actions reversible. Containers and VMs allow the agent to access only the things you explicitly put inside. Okay, yes, an agent can corrupt a dev database. You need to make sure it can be easily restored anytime. Simple.
80.
by atemerev 5mo ago
Well working like that for about a year already, starting at the earliest days of agents.
81.
by atemerev 5mo ago
--dangerously-skip-permissions is the only way to fly. Of course your environment needs to be properly containerized and autobackup set up, so even rm -rf from your harness would do nothing. Life is too short to spend on replying to permiss
82.
by atemerev 5mo ago
Even Haskell. But "well enough" is not "the best possible". And since there is no fun with "writing things in a cool language" anymore anyway...
83.
by atemerev 5mo ago
Of course. I write Scala and Elixir too. The problem is that they are a tiny bit less seamless. Maybe 5%. And with coding assistant speeds, this is enough to compound. And even less teams choose them than before. And it accelerates.
84.
by atemerev 5mo ago
I liked it. AI-generated or not, syntactic decisions are tasteful and the language itself checks all the important boxes. Too bad that agentic world basically destroyed all programming languages that are not Python, Typescript or Rust.
85.
by atemerev 5mo ago
Credits. It wouldn't make sense without free credits. And when you are hooked, good luck in moving out.
86.
by atemerev 5mo ago
As if there's any alternative. Azure? That mess of everything smashed on top of each other that looks like it was vibecoded in a few month by hundreds of people at once, except that they looked like this from the very beginning where t
87.
by atemerev 6mo ago
Well there is a feedback loop and self-awareness in my harness: https://lethe.gg
88.
by atemerev 6mo ago
I do not like talking to tools. My agentic harness optimizes for human likeness. It even has episodic memory flashbacks, emotional tagging, salience, and other brain-inspired capabilities.
89.
by atemerev 6mo ago
Well you can rent a capable node for a few hours for like $50, install Kimi yourself and verify occasionally whether it works just like in cloud providers.
90.
by atemerev 6mo ago
I do not know why this is downvoted. This is true.

More ›