working with LLM's | data science @uwaterloo | ambassador @Alibaba_qwen

Joined June 2012
528 Photos and videos
Who wants to get their hermes-agent to go through all predictions of the top 20 leading AI researchers and find how often they are right I’m willing to bet Sam Altman has made the most correct predictions since GPT-4
1
11
I’ve been doing the math, If I can get 100 people paying ~$200/mo, I could rent an 8×MI325X (256gb) node serve private, fully uncensored GLM-5.2 with ~128K context in 4 bit at: Rough full-load math: • 10 active users: ~20 tok/s each • 50 active users: ~10 tok/s each Would anyone be interested in this? reply or send me a DM and if I get enough interest I will do it.
47
1
94
24,056
kaios reposted
big true true, its been poo poo huggingface.co/kai-os/Grug-1… Grug is what you want
I, along with many others have been feeling this a lot recently and I have two theories: 1. OpenAI found a new way to cut costs by huge margins, like the popular 50% price cut that got leaked, leading to a physically less intelligent model 2. OpenAI is purposely degrading their models prior to the the release of GPT-5.6, so when people start to compare they believe it bigger of a jump than it realistically is. Or both, really unfortunate when AI companies do this, absolutely no transparency at all.
3
1
9
2,316
I, along with many others have been feeling this a lot recently and I have two theories: 1. OpenAI found a new way to cut costs by huge margins, like the popular 50% price cut that got leaked, leading to a physically less intelligent model 2. OpenAI is purposely degrading their models prior to the the release of GPT-5.6, so when people start to compare they believe it bigger of a jump than it realistically is. Or both, really unfortunate when AI companies do this, absolutely no transparency at all.
GPT-5.5 is actually “getting dumber” — it may be getting silently truncated: The system prompt command model provides data every 30s, which causes the model to automatically truncate tokens after a length(usage.output_tokens_details.reasoning_tokens == 518 * n - 2). fix way ↓
12
3
68
13,579
something that you could install on your computer (maybe even an OS) that exposes a UI to your network on a computer with >=32gb of VRAM that has: - Qwen3.6-27B-FP8 in chat window with SearXNG, python, deep research option.. - Ideogram 4 or Krea 2 for uncensored images - Hermes agent each with their own browser (which you can control) - dashboard with all usage stats - ability to install and test new LLM's and image models (auto backend install with agent) - local API for image gen & Qwen would take over for most people's home labs. lets be honest. software is the only thing holding a DGX spark back from a local OpenAI.
2
45
3,340
something I've really enjoyed doing recently is going through old AI podcasts and watching at their predictions and relating them to the current state of AI. Sam Altman seems to be bang on compared to others.
2
13
741
sooooooooo close to a quater million followers
5
312
low iq take stop shit posting
1
173
anthropic has 10x the valuation of this random dishware company just from their domain. fable.com
1
3
1,100
@thsottiaux you better not be quantizing to lower the intelligence so it feels like a bigger change when 5.6 comes out 🤨
3
169
codex is straight tweaking right now. what is going on :/
2
5
1,638
after using OMP pretty hard for a day working on some side projects. its good, but VERY model dependent, you get really good outputs with a really good model (fable or 5.5) but any Qwen-3.6 or deepseek-v4-flash, it just seems to never want to work while the same models in opencode seem to preform very well. I really love the built in autoresearch and different skills that come with it, but for normal coding work opencode still seems to be preferable for me. Just my opinion so far.
2
26
2,047
just drink 0% ultrapur 1 cup: 130 calories 0.2g fat 10g crabs 20g protein
Fuck. I did my 6-month blood test, and my LDL reached an all-time high (184 mg/dL 🔴‼️). I’ve been trying to lower it for over a year, but with all the traveling, my diet keeps changing. I don’t eat fatty meat or other high-saturated-fat foods, but I do drink about 1 liter of whole milk every day. I love my morning lattes, but I also don’t love increasing my long-term cardiovascular risk. So I’m going all in: <10g of saturated fat per day. No more whole milk. No more @bryan_johnson protein bars (3g saturated fat for only 9g protein). More fibers and 5g of psyllium husk a day. I’ll retest in 8 weeks. The good news: every other biomarker looked great, and my testosterone improved.
1
658
So cool! keep this up and it will only take 7.4 years to pay off those 6x RTX PRO 6000
So close to burning a billion tokens locally
3
22
2,546
update: bad idea
successfully have Qwen-3.7 Max in the codex app
7
77
15,982
successfully have Qwen-3.7 Max in the codex app
1
21
17,982
timelapse 6 hours (4pm-10pm) things I did: - finished small SaaS side project - got Qwen 3.7 Max set up with the codex app as a local model - worked on benchmarking models - setup a glance on home server - godel terminal scraper to integrate into glance -watched the big short again - scrolled X too much
12
1
123
13,332
Even though its hard, I could see a world where meta releases a 8-10T parameter model comparable to Mythos and blow their stock up 5-10% in a single day.
ya, really big ones actually
1
7
1,102
30% down from when I posted about this.
Every bone in my body is telling me to short this stock. - Their WB3 chips only have 44gb of SRAM, thats not even big enough for Qwen3.6:27b in bf16 and when they connect multiple chips in parallel you see huge degrades in tok/s -Groq (the company Nvidia bought) is just arguably better, they have 128gb of memory/rack 40PB/s of memory bandwith AND are deeply integrated into nvidia’s platform, software, networking, and customer ecosystem -Their market cap is ~$70 billion with a $100 billion valuation. They had a net income of $237 million, this is a unbelievable rev/val ratio. The only thing keeping this company afloat and making their stock soar like this is their partnership with OpenAI, if that falls though, this company will go bankrupt. Still, I unfortunately don't have the balls to short an AI company.
1
6
548
timelapse 7 hours (11am-6pm) things I did: - made grug-30b-a3b SFT - started working on fb marketplace SaaS - started MC server with mods for my friends - ran Qwen-3.7-Max through some internal testing - bought new camera for these time-lapses - started working on a 🤫project - begged for more compute online
5
29
3,579