@profondob1ue retweeted
Brev was acquired by NVIDIA 2.5 yrs ago. Going from startup to a big company was a big change.
Over drinks, I told Jensen:
“This is the biggest company I’ve ever worked for”
He smiled, looked at me, and replies:
“Me too”
I survived this long at NVIDIA because Jensen made it feel like working at a startup. He understood that layers of a company, necessary for operationalizing work, also suppress information flow. It happens naturally:
• People want praise, so they highlight the work they’re doing well & hide problems
• People want ownership, so they suppress ideas that don’t come from their org
But at Jensen’s Nvidia you could email 1000+ people at once to force information to flow outside of an org that might be suppressing it. In fact, you are expected to, sending an email with the “Top 5 Things” you’re working on to everyone at once, to force information to flow outside of a bad manager or org who might be suppressing it. I jokingly call email at NVIDIA a “mosh pit”.
He was always chasing information that was stuck in the corners of the company, always learning from the company he was leading.
NVIDIA is the world’s largest startup because there’s a founder driving it. Taking NVIDIA from its founder is a mistake.
“Find something that feels like work to others but feels like play to you” - @naval
Still one of the best life advices I’ve read so far.
1
65
“The machine that builds the machine”
@elonmusk gives more importance to the design of the factory to make a good Tesla more than the design of the Tesla itself.
I feel we are at this stage in Software as well, as engineers we should not be focusing on the software anymore but on the systems we have in place to make software.
A lot of people are still focused on how fast they can build the machine now, but they are still in the loop.
In Big Tech I’ve not seen many abstracting themselves away and building the machine that builds the machine, but maybe that’s just self-preservation instinct.
m.youtube.com/watch?v=f9uveu…
26
Can an LLM write neural network weights from scratch?
Not train one. Not fine tune one. Write every weight itself, from nothing.
Researchers at MIT CSAIL just showed it can.
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
1
60
Is this a new era for Distillation?
“Bro distill all your capabilities around XYZ into this smaller model, make no mistake”
rikiyat.github.io/blog/weigh…
39
lnkd.in/p/gFr3dhrs
Boomers gotta wake up.
18
Unexpectedly impressed by iPhone Duo.
Bought it with the idea to try and return it in a week, I know find is the perfect machine for the AI age.
You can peak at your favorite agent running all the time while using the whole for other mundane tasks.
It’s a first gen but great job on this Apple.
More people will switch.
1
85
We live in strange times.
Meta is now using the same marketing strategies of YC startups.
Big Tech capital + real startup mindset is the most dangerous combination in tech IMHO.
Bullish on $META
1
79
Impressive, in the end influence to invest in top companies was much more worth it than product revenues.
Attention is the currency.
💰 Passed $10M/y in revenue + investment gains this year
I've always been aggressively increasing my profit margins to as high as possible, especially the last few years, and especially this year where I've reduced my SaaS dependencies to just a handful of companies by vibecoding my own replacements
It's now about 94.5% profit!
Spend less -> lower costs -> more profit -> more money saved -> more money to invest -> more % returns on investments
And at some point that investment return starts to pass by your business revenue which it has for me for about a year now
Most of my investments have been simple ETFs like Vanguard S&P500, but also a few lucky stock and startup bets that worked out:
Back in 2022, I couldn't get any available GPUs, so I thought it might be smart to buy Nvidia if there's such a shortage and that was a good bet too
Then back in 2023 @dannypostma recommended me to invest in the companies we used for our AI inference for our photo AI apps, so I did that too. First I invested in @replicate and then also @FAL
From then I just asked every app/service I used if I could invest, like @cursor_ai in 2025 which then got acquired by @spacex recently!
My business revenue has remained very stable at about $200K-$250K/mo, but even with that investment returns have now surpassed that
Investments fluctuate a lot so some years it's up some years it might be down a lot (who isn't expecting a big crash at some point?), and most investment gains are unrealized (I never sell!), so keep that with a big grain of salt!
I keep my revenue/gains together with my profit margin and my sleep score in my situation monitor dashboard, as those are equally important :D
Many people in the last decade told me to spend all my money, either personally or with my business to grow it, we don't know how that would have ended up, maybe my companies would be bigger, but I'm happy I did it in my own lean profitable healthy way and it ended up great I think! 🙏😊
1
72
TBH I think it’s fair if we end up with a model that is similar to Formula 1.
In AI F1 different teams participate with their best models, a defined set of rules and judges to determine which company has the best frontier model at any given time.
The company winning the competition can re-use part of the technology developed in the AI F1 to transfer on their commercial model.
This will command a premium on their commercial models as they have shown to be the best in the AI F1 and have brand value associated with it.
As in F1 those companies will never release the exact models that won AI F1 for public access, for the same reasons F1 teams do not make F1 cars available to buy on public roads.
TBH I’m fine with that, despite what tech X believes we don’t need to risk human extinction just to slop our SaaS faster, exactly as we don’t need to commute in F1 cars just to save a few minutes on our daily commutes.
Understand that "pace the frontier" means "pace public access to the frontier" and it takes on a much darker implication
64
Latest AI leaderboard is dominated by 4 U.S. companies.
Two foundational labs, two big techs (Meta, XAI)
The only big tech still competing on the AI frontier are founder-led companies (Zuck, Elon).
This is no coincidence.
Google is reportedly slowing down the LLM foundational research, which caused the departure of Jeff Dean and others.
Every other “rational” big tech CEO (non-founder) is facing pressure to justify the impact of burning compute for research, and does not have the “political” power to just keep pursuing it.
Hence they are taking the “safe” choice of pull back.
I wonder, is this really the safest choice?
@ArtificialAnlys
1
94
@profondob1ue retweeted
Added Gemini 3.8 flash to threejseval.com and is doing a great job, comparable or better than Gemini 3.7 at @threejs!
@GoogleAIStudio @GeminiApp @OfficialLoganK at threejs evals we are trying to find the best LLM models for threejs. We do that with human pair blind voting. Go visit the website and check the update ranking threejseval.com/ranking
Added muse spark 1.3 free to threejseval.com and is doing a great job for the price (free) compared to any other model at @threejs!
@alexandr_wang @AIatMeta At threejs evals we are trying to find the best LLM models for threejs. We do that with human pair blind voting. Go visit the website and check the update ranking threejseval.com/ranking
@profondob1ue retweeted
Added muse spark 1.3 free to threejseval.com and is doing a great job for the price (free) compared to any other model at @threejs!
@alexandr_wang @AIatMeta At threejs evals we are trying to find the best LLM models for threejs. We do that with human pair blind voting. Go visit the website and check the update ranking threejseval.com/ranking
Added Fable 5.1 to threejseval.com and is on another league compared to any other model in @threejs evals.
No comparison really with the others. Blew up the rankings.
Also really expensive! it spent all the tokens i had left.
@profondob1ue retweeted
Form some early evals grok 4.6 high is better than 4.5 high in @threejs .
Here some preview images. not yet released.
Son on threejseval.com
4.5 High 4.6 High
LFG Italy 🚀 if you are building something interesting in Italy, let’s connect!
Italy now has 12,000+ innovative startups.
That’s almost 3x compared to 2015.
And this time, the timing feels different.
AI is changing what it means to build. You don’t need a 30-person team to ship anymore. Two people with the right tools can build, launch, and iterate on real products.
And the talent pool is deeper than most people realize.
Polimi, Bocconi, Sapienza, and universities across Italy are producing incredible builders.
And many others are coming from outside the traditional university path entirely.
The opportunity is here.
Now we need to build.
1
2
545






















