Harrison Kinsley

I do not miss claude at all

@yacineMTB Use a mic thats easy to use. I love the dji mics. No more in your way. Whatever 1080p cam. Ppl care about the content. Some channels benefit from this stuff. If youre trying to zombify ur audience then u need it to be beautiful and gr8 audio.

Day 2 of being a professional YouTuber. I just spent 5 grand on camera equipment and recording devices and lighting fixtures after watching tutorials that were secretly ads on how to become a good videographer. It all arrives tomorrow

can't wait to be fighting the narrative that mythos "hacked the nsa" for the next decade.

@DanielWhit21874 Glm 5.2

@Sentdex What are you using ?

I do not miss claude at all

@OlympusBets I dont think fable was that good. Ppl just want what they cant have. Wont buy a sub again for fable even if they make it avail again.

@Sentdex Don’t lie, you still want Fable 5

@fourweekmba Its true. I really don't miss it at all.

@Sentdex None believes that

@tabbislashcat Do it!

@Sentdex my paw has been wavering over the cancel button all day..

@AndreBuckingham Damn

@Sentdex lmao... not really, just playing along with your post... i use all the models... "dang" is a crazy word though... just say "damn"... it´s cleaner... gets the message across better

ur right we should trust the sycophantic models that are taught to purposefully mislead us, from the companies doing their best to stop us from being allowed to do matmuls

@MosesSpacely the open source part. thx for this easy layup wtf.

@Sentdex How can you trust a Chinese open-source model?

@AndreBuckingham dang i just met my first actual claude simp thats crazy.

@Sentdex that´s ok, claude doesn´t miss you either

@benteisheuer subsidization is fake. There's no such thing. it truly does NOT exist. The reason local compute is so much $$$ is bc people buy cheap "subsidized" subs from these companies, then they raise $$$ on inference demand growth and buy up all the hardware. Cycle continues.

@Sentdex GLM doesn't subsidize as nicely though

@hellomhnd I'm using it full time for real work. Public project i'm using it for that you can see now though is: github.com/Sentdex/minion

@Sentdex Is glm that good? Did you build any public projects with it?

@HaseebMir91 @dcbuilder @UnslothAI @togethercompute hmmmmmmm. No i dont think its that much money even in today's inflated prices. If you pay $80K for 4x rtx pro 6k u definitely messed up big time lol. I can imagine some 3rd parties charging $60K, but you can build it yourself for cheaper than that even at today's inflation

@Sentdex @dcbuilder @UnslothAI @togethercompute How much this setup costs now? Grok: The local setup Sentdex described (4x RTX Pro 6000 GPUs running GLM-5.2 quantized via Unsloth + llama.cpp) currently costs roughly $60,000–$80,000+ USD for a complete working system as of June 2026.

@HaseebMir91 GLM 5.2 IS accessible via API from ~ a bazillion providers. I would recommend @togethercompute

@Sentdex You have GLM 5.2 Locally lol bro

@weywadt lately, a lot of robotics, reinforcement learning, and simulator work. also been using it to build a coding agent: github.com/Sentdex/minion i friggin love this model

@Sentdex what do you use glm with for coding?

@mSanterre @LLMJunky @datacurve @winkey_h It's better than Opus 4.8 and GPT 5.5. DeepSWE is a great benchmark, but benchmarks are never a perfect picture.

@LLMJunky @datacurve @winkey_h did anyone say it was better than opus 4.8? I feel like it's better than 4.7 and on par with 4.6.

GLM 5.2 is now on DeepSWE as the top open-source model on our leaderboard. With a pass@1 score of 44% at max effort, GLM 5.2 is indisputable #1 open-source model besting Kimi K2.7 Code by 17%.

@LLMJunky @datacurve @winkey_h It's better than Opus 4.8 and GPT 5.5. I am not in shambles by a benchmark not agreeing with my assessment. I am actually excited to see how strong it performed here to further confirm what I felt.

@datacurve @winkey_h The "better than Opus 4.8" people are in shambles right now

@JsonBasedman @djpardis seems to be a bunch of fans of pi! I need to start a beef with them on X for marketing for minion, the best coding agent there is

@Sentdex @djpardis Dynamic workflows are great, and it turns out Pi has a good extension for them. So yeah leaving Claude code and not looking back.

CC is slow and lacks features. I'm only using it now to maxx out the credits I paid for

@kristoph @IurySza interestingly, GLM 5.2 seeeeeeeeems to not reason much, if at all, on simple tasks. Sometimes it does just go straight to output. I'm not sure it's ever worth actually changing off max.

@Sentdex @IurySza for simple tasks you can run with reasoning node and it will still perform reasonably well.

He's running glm 5.2 at decent speeds locally. IMO anything above 20 is decent, aber 40 is quite OK. Glm is kinda token hungry though, so this still may not be where it's actually usable though. Maybe in 6months.