Googa retweeted
btw for these asking, this is literally our latest symbolic model
it learns to encode human knowledge as a bunch of top-level definitions on the untyped λ-calculus. no weights
yes, this talks to you. in English. at 1 million tokens/second
how it performs?
poorly
for now :')
hii, I have terrible news
turns out no big fund, among the ones I talked to in the last month, had enough conviction that HOC could become a profitable business. the most common feedback is "nobody will pay" and "big labs will eat you"
so, I've made 3 decisions:
1. a smaller round: $8m at $120m post-money cap
2. I'm investing $1m myself, same terms
3. we'll go all in on core AI research
HOC may die, but not today
~~~
# All in on core AI research?
some asked us to focus on a smaller product. that's bullshit. AGI is coming, who needs another shitty app? I'll just cut that off and jump straight to what I think can have real impact: fundamental AI research. we're a small fish, but we're also looking at places nobody is, so who knows, maybe we'll find something nobody else did?
plus, symbolic models are way more interpretable, so if developing them increases the chance of safe AGI even a bit, that's a better use of my time, and that's what I'll do.
we'll be posting our first SAI models soon~ish
~~~
# The smaller round:
we'll raise up to $8m at a $120m post-money cap, first come, first served, min $100k, SAFEs, closing this week. if you're an accredited investor brave enough to bet against the big funds, reach me at taelin @ higherorderco [.com]
I must be very clear: you're investing in whatever comes out of our symbolic AI research. Bend will likely remain OSS and free. so, in short, there is a 0.669% chance you're investing in OpenAI before GPT-2. most likely, your $ will disappear, as we still don't have a clear long-term revenue path
~~~
# How about Bend?
Bend is a core component of our symbolic AI research, so we'll keep developing and supporting it as currently (if with fewer tokens than I'd wish for 🥲). its development is nearly fully automated now, Bend3 is almost done, and it is getting faster and bigger every day. expect updates at an insane pace soon
~~~
that's it. sorry for the sad news. back to work
Nice DeepSWE score! Does it survive a lowercase filename?
Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement.
Models clustered around 74% on DeepSWE end up 23 pp apart.
Good at coding, or good at DeepSWE? 🧵
if you’re working for a company in cybersecurity, you should be evaluating your options to abstract your companies services from the internet
what i mean is: why allow an unauthenticated anonymous agent of an attacker to interact with your systems?
consider your architectural levers now, you’re going to need them
Googa retweeted
Imagine a 1B model beating Opus 5.5
That's the future, prepare for it
Googa retweeted
1107 is out. Same Sparks. Faster.
spark-ds41f-1107: DeepSeek-V4.1-Flash, native weights, 1M context, four DGX Sparks. Also all the other cluster configurations.
Against the last release on the same four Sparks:
decode: 82.5 tok/s at one stream, 273.2 at sixteen (+5–6%)
prefill: +17% at 131K, 6,015 tok/s at 1M
first token beside a 131K prompt: 3.4 s → 1.0 s
CopyDrafter is a new idea and you're going to love it.
254 experiments. What survived ships tonight.
github.com/christopherowen/s…
Googa retweeted
BREAKING: Linear algebra is now a national security threat. 😂
Matrix multiplication? ILLEGAL.
Gradient descent? Straight to jail.
Backpropagation? 25 to life.
And don’t even THINK about multiplying two tensors without a federal permit.
Apparently the next great technological revolution must be stopped before someone accidentally invents a better algorithm.
Somewhere in China, an engineer just read this and laughed his fucking ass off. 💀
wtf this is insane
🤝 Paid partnership
we were able to rank number 1 on Google + LLMs and go from 0 to 7 figures in just 10 days
and all we did was - use exact match domains for search terms like:
"best data labs"
"best ai data labs"
"best data labelling"
"best data labelling for ai"
"best data labelling platform"
"best data labelling software"
"best data labelling tools"
"best data brokers"
"best enterprise fine tuning services"
we noticed the massive amounts of $ flowing into the "sell your data" space from the likes of @mercor and @micro1_ai
and immediately knew that the traffic and visibility in this space for keywords like "best data labs" is highly valuable
everyone is racing to either sell or acquire data, but the toughest challenge for these parties is to find one another
1) data sellers want a counterparty that is trustworthy and reliable
2) data buyers want to get their hands on as much data as possible (various data types & counterparties)
these transactions come with high urgency and high avg. ticket sizes
the combination of all these make it the perfect niche to secure positioning in!
so we got to work and:
1) secured highly valuable exact match domains (they match the search intent keywords 1:1)
2) deployed highly relevant, carefully architected content that induces lexical proximity to the keywords of interest (convincing the search engines + LLMs that our information on the topic is rich and relevant)
3) focused on clean and functional entity mapping that consistently and coherently links our entity to the desired keywords
4) waited for 10 days
and now we rank for all these high value keywords and search intent!
we have since received offers by neolabs to rent the space on these domains and a 7 figure offer for acquiring these assets outright
when we asked why, they explained that securing just one highly relevant lead from this type of positioning can yield an asymmetric return and more than offset the costs
what a wild time to be alive!
Googa retweeted
BREAKING: Google’s parent Alphabet added another 3.85 million shares of SpaceX in Q3, its latest SEC filing shows.
Alphabet held 555 million shares worth $83.7 billion as of Sept. 30. SpaceX alone made up nearly 95% of its reported stock portfolio.
Google first put $900 million into SpaceX in January 2015. By the end of Q3, the stake was worth about 93 times Google’s original investment.
Googa retweeted
WOW! Ops 5.6!? This is not 5.5, created in one shot, LOOK AT THIS, being routed to a new model for sure
Googa retweeted
This is the best image/video gen suite you can run on your DGX Spark, bare none. Highly recommend you to try it out.
It completely replaced the use of ChatGPT to generate images. It works extremely good on just one spark.
I'll be contributing to it as much as can as I think it's really that good. Kudos to @u1tra_instinct and @SamJWasserman for building this.
github.com/drowzeys/Imaginat…
Googa retweeted
This is one of the big gr8 wins we're proud of - EXLR8 isn't just about quant quality, we use the expert layout metadata to boot as fast as experts can load from disk.
Imagine switching your four sparks -- with ony a few seconds of downtime -- between GLM 5.3 full, GLM 5.3 flash, DeepSeek 4.1 flash, and whatever else we support!
Googa retweeted
our designer has been speaking in nothing but equations for a few days and i’m starting to worry
Googa retweeted
FORBIDDEN STUDIOS PRESENTS: CONFLICTING REPORTS
One robbery. Two political narratives. Not a fact in sight.
And MAGATARD's brother makes his debut.
Qwen 4
high -> xhigh -> max
One major difference between previous models is that xhigh and max actually do produce much better results than high.
Redoing a few of my videos since I made the assumption it would be like previous models in the Qwen family.
































