Cuth (@ItsCuthulhu) header
Cuth (@ItsCuthulhu) profile picture

tinkering | Analytics Grad & Analytics Lead @georgiatech

Philadelphia, PA
Joined March 2026
Photo by Cuth (@ItsCuthulhu)Photo by Cuth (@ItsCuthulhu)Photo by Cuth (@ItsCuthulhu)Photo by Cuth (@ItsCuthulhu)Photo by Cuth (@ItsCuthulhu)Photo by Cuth (@ItsCuthulhu)Photo by Cuth (@ItsCuthulhu)Photo by Cuth (@ItsCuthulhu)Photo by Cuth (@ItsCuthulhu)Photo by Cuth (@ItsCuthulhu)
NEW SOTA DECISION MODEL: SubQ Snap SubQ Snap absolutely crushed the S1 Bench, beating Jev!! Insane. 1,999 decisions, 0 errors. Macro accuracy 0.825, the best full run the bench has ever recorded. Jev scores 0.775. Mercury Decide, the previous best API model, scores 0.802. Subset scores: paws 0.940, vitaminc 0.917, boolq 0.910, massive 0.909, aegis2 0.856. Only helpsteer2 lags at 0.416, and that subset drags everyone down. To be fair, none of the other decision models do well on helpsteer2... And it seems to have done better than Jev. Subquadratic claimed 61.7 vs Jev's 57.9 on their own Decision Index. On an independent battery it is not even that close. It beat its own benchmarks by a lot. It didn't hit the speed benchmark on my bench but the accuracy more than makes up for it. About a second per decision, and the entire 1,999-question battery cost roughly 3 cents of input tokens, output free. To be fair, I felt it was the fastest but my agent said it was a bit slower. Nevertheless, this is the SOTA decision model. Hosted API only and still a preview model, so the open questions are weights and what the 384K context does on long inputs. But as a decision model, this is the new bar. This is my new go to decision model. Full results: bench.jakecuth.com subq.ai
4
5
3
33
7,222
Get ready
4
172
Cuth retweeted
BREAKING: Jensen Huang reportedly painted a fake tunnel on a brick wall, tricking an employee into running headfirst into it.
280
1,242
113
21,506
377,084
“We got this” is gonna go down in history
did the NVIDIA board just destroy $2,000,000,000,000 of value...?
42
We got this
BREAKING: Nvidia loses nearly $500,000,000,000.00 after SEC announces investigation into its accounting practices 10% chance Jensen returns as CEO.
26
It's time to buy folks. These are the typical 2 year returns after pullbacks, corrections, bear markets and crashes. Start deploying that cash.
48
JUST IN: 🇺🇸 President Trump calls for Nvidia $NVDA to reinstate Jensen Huang as CEO.
385
1,031
144
16,529
551,459
Nvidia board taking the entire Nasdaq down. Idiots of the highest order. Honestly, a site to behold.
3
169
Told Fable 5.2 to make a James Bond intro. It took 10 hours to make this and burned a month of tokens on @CommandCodeAI Some parts are actually pretty cool and some are laughably bad. It was just given references and set on /goal. I honestly think CommandCode CLI is just shitty.
46
Buy when everybody else is fearful. Works every time.
Trying to find the best price to buy Nvidia leaps and more Anthropic shares. Most people can't beat SPY cause they're ruled by fear and greed. If you learn how to read financial statements then this is generational buying time.
30
Trying to find the best price to buy Nvidia leaps and more Anthropic shares. Most people can't beat SPY cause they're ruled by fear and greed. If you learn how to read financial statements then this is generational buying time.
1
76
THE BANKS ALL WROTE THE SAME 2027 OUTLOOK: JPMorgan, Goldman, Morgan Stanley, Barclays. Different logos, one report. AI is still the engine: $1.5 to 1.8 trillion of capex next year. But the question flipped. It is no longer who can build fastest. It is who actually gets paid. The rest rhymes too. Growth cools but holds. Inflation cools, yields do not. S&P targets near 8,700 built on earnings, not hype. And everyone quietly diversifying out of the same ten stocks. When every bank agrees, what is the trade they are all missing?
85
Buying all these companies for pennies
49
Cuth retweeted
I agree with this assessment. If your goal is to achieve outcomes and you have both creativity and the ability to communicate, you're sitting on a golden age of earning capability.
My “is engineering a valuable pursuit” math. An outcome-oriented, creative engineer is now a 1000x engineer. A tabs-vs-spaces-debating, miss-forest-for-trees engineer is now a -10x engineer. The regular knowledge worker is now a 2x engineer. This is also backed by the distribution of token consumption we see in both our startups and enterprise base. Engineers and software factories will consume exponentially more tokens. If there’s a takeaway from Anthropic’s success, it’s that coding is the #1 killer app of AI.
1
3
54
Time to buy
JUST IN: Stock market return to "extreme fear"
1
150
Love rockstar but nobody considers that anthropic got to $30 b revenue in 1 year.
JUST IN: GTA VI surpasses $2,000,000,000.00 in revenue
39
But I don’t understand growth 😔
The only chart I care about for the IPO. Anthropic made about 12x more money from Q3 2025 to Q3 2026 than all other SaaS, combined.
2
182
Cuth retweeted
Introducing Da7em Bench v2.0. A fully independent benchmark for AI models, built on real client work. The first of its kind in the world. How it works I test each model across a wide range of scenarios. I compare its decisions with the choices people make in the same situations and with what people generally agree is the most sensible choice. Models still fall behind humans in several areas. A higher score means the model’s decisions are closer to human choices. A lower score means they are further away. As before, I test each model impartially across multiple harnesses. The standard is professional work. I judge each result against what a paid professional would deliver for the same brief. Version 2 covers 11 models across 12 axes. Each axis is scored out of 10, and the overall score is the average of all 12. The tasks and evaluation methods stay private so they don’t leak into training data and affect future results. Da7em Bench helps you choose a model for each task. The model with the best visual taste may not be the best writer. The model that plans best may not be the best at carrying out that plan. Da7em Bench is fully independent. No sponsors or outside funding. I’ve paid for everything out of my own pocket, thousands of dollars so far. There are many more models I’d like to include. This is a personal benchmark, and covering them all on my own takes time and work. The four images show the overall ranking, the scores for every model on every axis, the twelve axes grouped into four pillars, and a side-by-side comparison of Claude Fable 5.2 and GPT 6.2 Astra. You can explore the full benchmark and compare any two models here. da7tech.com/da7em-bench/?lan…
61
30
5
235
22,465
About 2 months ago, @typesafeai released Jev which was the first 'Decision model' to go viral. Some may say 'Laya' or another model came first, but it matters more which one starts the wave imo. Since then, I've been benchmarking, ranking and scoring these models on my DGX Spark for one reason... I wanted to use one for work but it needed to be the best one for my purposes and not collect data like @typesafeai. For a while, I've had one design style for the site and only basic charts because my agents and DGX Spark benchmarked them in the same manner. Now, I've added more charts and filters so you can determine which model works for your purposes. I appreciate the support!
5
1
3
618
Lastly, I added a Pareto Frontier to show size vs accuracy. This is a good way to quickly determine what you can run locally... Especially if you cannot have another company collect your data.
20
This chart allows you to see which are cheaper and includes local models as well. pplx-decider-v1.2 is the highest ranked for speed and accuracy, and costs nothing if you have the machine to run it. That being said, most cannot. If you cannot, then I recommend either using SubQ Snap 1.2 OR if you have 16+ GB of RAM use pplx-decider-flash-v1.3. You may also want to use an image decision model. I use these to classify social media posts for my job. It can watch a 2 minute clip, snap a screenshot of each second and break down what the content is about. This is way undervalued for data scientists. If you have questions about which system one (decision model) to use... I am your expert.
1
64
This chart allows you to figure out which models are both faster and more accurate than Jev. This does not make them cheaper. For that, you may want to click into it or read the stats in the table below.
52