Virtual Kenji⚡️

Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.

@karpathy Man I might actually need to renew my anthropic solution now Next Anthropic release: 1-shot Karpathy Wiki LM installation from within Claude

stop what you're doing… here's 14 things you gotta do with AI: 1. export your ChatGPT + Claude memories, context, and workflows so you know what each model knows about you that the other doesn't 2. generate DESIGN .md + CLAUDE .md + AGENTS .md for your agentic repos 3. build a model agnostic skills library that Claude Code, Codex, OpenClaw, Hermes, and Cursor can all read from the same symlink 4. version your system prompts in git with PR review so silent drift gets caught in diff instead of in production 5. build a /goal template library covering your top 10 recurring workflows so you stop rewriting the same 200 word prompts every Monday 6. set up a daily report and a weekly standup with your agent that reads yesterday's commits, tasks, memory, your calendar, and your inbox, then briefs you in 60 seconds 7. wire your Obsidian vault as the read/write target for every agent so research, decisions, and meeting notes land as linked markdown by default 8. build a Karpathy-style autoresearch loop where your agent logs its own failures, proposes skill patches, and you approve them weekly 9. design a UGC AI character (and/or mascot) with 35+ emotions for brand world building so every loading state, every error screen, and every post in your feed shares the same look 10. cancel 2-3 SaaS subscriptions by replacing them with agents and skill files 11. run a security agent across all your repos weekly to catch npm attack vectors, prompt injection, leaked .env, and MCP misconfigs before someone else does and you regret it 12. build a prompts library of image, video, and audio generation templates so you stop relearning Nano Banana, Grok, Seedance, and ChatGPT syntax every time there’s a new release 13. build a personal benchmark of 20 tasks from your real work so you can A/B every model release against your actual workflow instead of the bogus leaderboards 14. build daily cron jobs for your data ingestion across the topics you actually care about: research, content, trading, competitive intel, etc. the next few months will separate the operators using complex systems from those still manually typing prompts one at a time. these are table stakes. act accordingly.

xAI has expanded access to X Premium+ subscribers in Hermes Agent. Enjoy!

this is amazing. do this. get codex to write its own /goal /goal + gpt-5.5 high + fast mode is the highest leverage ai agent configuration available today

tips for codex goals sure you can use /goal but it also has a set_goal() function its almost better to prompt the model to set its own goal, it will likely write a better prompt than you

it costs less for claude to hire a dedicated copywriter than it would for one of their senior marketers to quality control LLM outputs… not to mention that the copywriter they hire will 100% use an LLM to help in the copywriting process don’t see why this is something controversial

spent a few weeks deep in Hermes + Openclaw earlier this year... over-offloaded everything, daily maintenance got brutal and output rarely hit switched back to Claude Code via Telegram and rebuilt the workflow from scratch then GPT-5.5 dropped... then GPT-Images-2... then GPT-Realtime-2 and Hermes quietly became THE move > best coding model > best daily model > best image model > best voice model all running through ONE subscription on your own agent takes 5 min to setup on a VPS, plug in your GPT sub, hand your friends and family an agent that generates awesome images, processes voice notes, builds anything and actually remembers everything this thing rarely misses anymore

hear me out anon. whatever you're using right now, drop it for one night. install hermes agent. set a /goal for that one thing you've been thinking about but never started. go to sleep. just try it once.

not sure if btc continues to provide diminishing returns but it is certainly possible (i choose to believe it wont for now) but alts (on average) will continue to get more difficult and dangerous to trade imo, irreversible trend. (1) massive competition amongst increasingly sophisticated buyers (who simultaneously believe less) (2) primarily traded on perps/with leverage. (3) huge dilution in coins with mkt valuing stuff high by default without justification (4) launch FDVs always capturing 100% of optimism for the asset without respecting price/valuarion (5) too much pre-market price discovery for (4) to be safe the average 2017 buyer buys spot and hodls weeks/months cos they believe, add on the way up, asset was trading at low val early so works out. average 2025 buyer is buying on perps without checking the valuation and sells whenever their PNL goes red or force sells in liquidation. however, it will remain the best place for returns for smart ppl. skill expression and asset selection is much more important. patience much more highly rewarded over being "early" on liquid markets last few years. and then outlier assets will continue to exist, maybe 1 every couple of years, and when they turn up you can turn brain off and 2017 it. imo anyways (hopefully)

had a fantastic time with Doc, VOD is available for replay if you guys wanna check it out will def have to do another one soon but in the meantime big thank you again to Doc and all of you for tuning in <3 see you bozos soon

@nellyeeee Listening to it now Insanely gud bro One of my fave paragon streams so far actually

@kloss_xyz Yup I actually went from $200 to $100 on Claude because of GPT 5.5 And now considering moving Claude down to $20

I’ve spent the last two weeks coding from the moment I wake up until the moment I can barely keep my eyes open. Codex with GPT-5.5 is that good. First AI model that’s made me seriously question whether I need $200 Claude. And I really don’t say that lightly. Anyone else?

current market thoughts: some alts straight up look macro bottomed, but that doesnt mean a bull market has to start today, they can take forever to build up enough steam for that, mini rallies like 2022 most likely. alts with narratives = tradable alts without = require crime, not worth your energy btc in 60ks is a steal, i believe in new ATHs so imo free sized spot 2x. take it with a grain of salt if you must- im a delusional btc believer. imo current btc rally most likely nearing the end soon as we trade into HTF downtrend structures. +28% in about 30 days. I was a bull from 60s and fighting off war fud, equities weakness etc but now imo the EV for bulls has diminished dramatically "if ES pulls back, then BTC dies" is low EV HTF signal, we more often then not front run equities weakness on HTF then stall out as they find a bottom. ES isn't a good leading signal in my experience. thoughts, not predictions.

compute is fucking king and gpt codex currently wears the crown, look at this thing: > guy tried to recreate Maplestory (played by 100Ms globally) with codex > 30 minutes in codex already one-shotted the design, gameplay and characters. > only used 5% of his 5 hour limit. do you understand how fucking insane that is? 2 months ago openai sucked at coding.... software engineers universally dismissed codex as a toy vs. claude code. Now not only is codex 5.5 the leading coding ai, it also thinks for longer, works 24/7 and is AVAILABLE TO YOU i love claude, i really want to try mythos but i can't. either thats because it's too dangerous or anthropic is still scaling compute either way - openai has the compute, they're resetting limits every other day and it's proven you can build insane things with it compute = abundance

I put everything into crypto for 3 years straight, only to walk away with 5% of what I made. The real cost though, is a mental one. The ones I care about who I could have, should have, helped. Instead I let greed get the better of me and in this game, you only have yourself to blame. What is only possible in crypto, is also only possible in crypto. Manage your greed or be punished by it a thousand times over. I share this in hopes to help just one person who’s currently at the other end of my situation, pushing for me. Be grateful for what you have and if you already have the life you dreamt of, secure it at all costs.

i'm done. codex is fucking incredible after heavily using claude code for over 13 months, i've moved to codex opus 4.7 is painfully slow and takes 5-10 mins for a one-liner. the app is super buggy and flickers constantly. low thinking is useless. and they keep nerfing the model for some reason?? codex's new app is genuinely beautiful and gpt-5.5 thinking-medium is the perfect balance ngl @sama you cooked on this one

@gkisokay Actually have a spare 3060 I will use this for exactly Thanks so much for this post bro

The Local LLM Cheat Sheet for Your 8GB VRAM or Unified-Memory Device If you have a spare base Mac mini, RTX 3060, or 8GB iGPU device, these are the top LLMs you can run on them. Best Daily Driving Models Qwen3.5 4B - The Unified 8GB Daily-Driver Newest 4B from Alibaba and the strongest one-model answer on 8GB. Q8 on a discrete GPU is essentially lossless, while the Q6 fallback for unified memory still beats most 7B Q4 quants on benchmarks. Qwen3.5-9B - The Upgrade Daily-Driver for 8GB The best overall model that fits VRAM. Q4_K_M is just enough, and Q5_K_S at 5.82GB is also pick-safe if you want to push the ceiling. Not a clean unified-memory fit. Phi-4-mini-instruct - The Best Unified-Memory Microsoft’s strongest small instruct model, and unusually clean at this tier because the same Q8_0 file fits both ceilings. Great for chat, summaries, structured output, and function calling. Gemma-4-E4B-it - Newest Top Model 4B effective active params, sharper instruction following than Gemma 3 4B, and best-in-class IFEval at this size. GPU users get a near-lossless Q5_K_M, but unified-memory users have to drop to Q3, which dulls reasoning. Best Reasoning Models Qwen3-4B-Thinking-2507 - Best Reasoner The best small reasoner at this tier. Q8 on both passes means the CoT doesn’t quantize-degrade. AIME and MATH scores land ahead of Phi-4-mini reasoning, making it the reasoning pick whether you’re on a GPU or unified memory. DeepSeek-R1-Distill-Qwen-7B - Best Chain-of-Thought The deepest chain-of-thought option if you can spare the VRAM. Distilled R1 at 7B is the strongest reasoning chain at this size, but there is no clean sub-4GB unified-memory quant that preserves the reasoning chain, so is GPU-only. Best Specialist Models Qwen2.5-Coder-7B-Instruct - Best Coder The best small coder for HumanEval and MBPP. Use bartowski’s GGUF repo, since Qwen’s official GGUFs have size anomalies and Unsloth doesn’t ship a 7B-Coder GGUF. Best for code completion, refactors, debugging, and repo work. Qwen3-VL-4B-Instruct - Best Vision It replaces Gemma 3 4B as the vision choice with same 4B-class footprint, newer post-training, and 2x the context window. Q8 on both tiers keeps both the language core and vision head near-lossless. SmolLM3-3B-128K - Best long-context 28K context at 3B params, lossless BF16 on GPU, and near-lossless Q8 on unified memory. Pick this when the job is feeding long documents to a tiny model that can keep up. On a 3060/4060 with VRAM to spare, Qwen3.5-9B is the best pick right now. The other rows only matter for specific jobs: Qwen3-VL-4B for vision, Qwen2.5-Coder-7B for coding on GPU, and SmolLM3-3B-128K for long-doc work. Which models have you been running on 8GB VRAM or unified-memory devices? Let me know in the comments.

A common dynamic I observe with AI: it feels most impressive when you don’t know much about the subject, don’t care or don’t have a clear idea of what the you want. This applies across design, code, legal, and more. If I don’t know code very well, every piece of code it writes feels very impressive. Once you know what something should feel or look like, it becomes almost impossible to guide AI there. And you definitely can’t one-shot it.