zach
@blip_tm
Followers
3K
Following
56K
Media
192
Statuses
2K
in 2024 i:. - sold a chip startup.- got hired to lead an AI chip design team.- moved to new york city.- hired 4 fantastic silicon engineers.- made a ton of new friends.- cooked feasts for said friends.
27
29
1K
going to tons of parties for 6 months to a year cures you of fomo for life.
You should degen-maxx in a controlled way when you're in your 20s so you don't crashout when you're in your 30s and actually ruin your life.
54
1K
20K
your startup is just an openai wrapper. openai is just an nvidia wrapper. nvidia is just a tsmc wrapper. tsmc is just an asml wrapper. asml is just a sand wrapper.
289
748
11K
i was depressed at 22, so i founded a deep tech startup and started dating a model and i immediately got better.
84
199
5K
@sneaknsneak Thereâs an entire body of research that shows that women have more fluid sexualities than men! Most straight women donât have a disgust reaction to the idea of lesbian sex, and the same is true for lesbian women and straight sex. This isnât the case for men. We donât know why!.
38
8
500
@BLUECOW009 please do not use a homegrown cryptographic algorithm for anything you plan to deploy publicly or handle real user data!. even the best cryptographers and engineers can build systems with flaws (look at Rainbow).
2
2
452
@NEETWorldOrder itâs really easy for her to say âbeing hot doesnât matterâ when sheâs hot.
2
2
338
startup pitch: itâs just a chip company but itâs run by me and my friends and weâre all really good at building chips and really good friends and we make chips together and hang out.
18
5
241
i know multiple startups who won over key talent with the promise of living in nyc rather than sf.
So many YC founders moving to NYC post batch and, when asked why, mention the social and cultural scene. NGMI. SF is less happening, but who cares? Thereâs (objectively) nowhere better to build, and you should be focused on your startupânot how late the bars stay open.
5
6
212
this'll probably be my most controversial blog post yet:. you probably shouldn't drop out of college, especially if you want to design chips. âŹď¸.
9
7
182
@BLUECOW009 what hardness assumptions are you relying on? and what's the enc/dec performance? i feel like massive keys would hurt performance significantly.
1
0
169
@albfresco @ctjlewis when you grow the crystals from molten silicon, theyâre in giant cylinders called boules. and itâs easiest to trim them into rectangles later, at the same time youâre dicing the wafers, rather than doing two steps of wafer cutting.
3
0
154
boss makes a dime, i make a dime. we work at a startup and most of our comp is equity.
5
1
130
@animalologist I think the vast majority of men would take this extremely poorly as well. There just happens to be a small contingent of men who are very self-conscious about their desirability who are also incredibly loud online.
3
5
114
@BLUECOW009 for cryptographic algorithms, performance is reported in cycles as they're all fairly fast. nearly all secure implementations should be constant-time (is yours?). though before you think about implementation security, you need a security proof.
1
0
117
@CartoonsHateHer i think this is the case for most non-traditionally-attractive physical features. as a guy who likes tall women and women with short hair, its easy to bat above my league because so many men see those as deal-breakers.
3
2
96
@dela3499 my rule of thumb if you had a date where you didnât vibe:. after the date, donât preemptively send a âno thanksâ message. if they donât text you first, itâs okay to mutually ghost. if they text you asking about a second date, politely say no rather than ghosting.
7
2
95
so many AI chip startups are allergic to HBM. cerebras, groq, and now d-matrix have huge performance drops once a model canât fit in on-chip SRAM. this means you either need insanely large deployments, or get poor performance. just use HBM, guys!.
9
1
85
the number of hot people you see in the average day walking around new york can absolutely one-shot an sf tech boy.
3
0
80
The Dilithium implementation in Google and Microsoft's Caliptra root of trust just got hacked. By measuring the switching power consumption of internal pipeline registers, attackers extracted keys with just 10,000 power traces âŹď¸
11
7
77
i work at normal computing now!. teams with genuinely new hardware ideas are rare. teams with the vision to tie them to real-world problems are rarer. normal is one of those few companies, and im super hyped to help bring this vision to life.
Proud to officially be working with @blip_tm. There are folks you meet and realize in minutes you would relish building them up, and deeply fear betting against. Zach is one of those people. Same for @thomasahle @mengdi_en @nmartyanoff and @zaqqwerty_ai and many others at Normal.
9
4
62
@GarrettPetersen "ready or not, here i come" is the only answer here, right? is there something else people say?.
1
0
60
most of the early AI chip companies are struggling because they didnât realize how big models were gonna get. that means the most important part of your chip isnât compute itâs memory hierarchy.
4
3
60
two months ago i moved across the country to a city ive never lived in. tonight i made homemade carne asada for 30 friends. thanks, guys :).
3
0
61
if you want practical open source hardware, spend your time lobbying TSMC and GF, not writing tools for outdated skywater pdks.
If I wasn't at a hardware company I would open my macbook air and work on Open source semicnodcutor chip design software Every Day.
8
1
54
selling silicon ip is a lot easier than selling eda tools. seems like a startup could build great in-house ai-powered design tools and use them to build better ip blocks for cheaper.
6
1
51
met a girl and she immediately asked how much money i made. is this how people flirt in new york.
8
0
50
in honor of hitting 1k followers, hereâs a real actual picture of TSMCâs PDK documentation
3
1
47
@kitten_beloved there are definitely some wealthy / successful men out there who really want successful partners⌠but I donât know how many of them are still single at 34.
2
0
47
itâs not neuromorphic computing if there are no spikes. otherwise itâs just a sparkling non-von-neumann architecture
4
3
46
proud to by the only AI chip designer capable of making something other than a systolic array.
4
0
45
hey anon. do you want to build chips with your friend zach?. apply here:
3
7
42
finally get to announce that my prev startup got acquired!. excited to see cryptographic processing-in-memory technology keep maturing đ
6
2
42
@HotForMoot took her on a road trip with every stop planned out to be something cute that referenced our relationship somehow . and I co-founded and run a semiconductor security company.
1
0
42
some exciting news: im joining @BTQ_Tech as an advisor!. i'm excited to work with the whole team at BTQ to develop state-of-the-art hardware and software for post-quantum cryptography. right now is a great time to be building quantum-safe hardware -- read my analysis below âŹď¸.
9
8
44
FPGAs might have a brief moment in the sun for AI acceleration, because none of the commercially available AI ASICs have HBM, but the fanciest FPGAs do.
5
0
44
Oscillator Ising Machines (OIMs) are a new chip architecture for solving optimization problems using physics. But theyâre only accessible to researchers at the few labs where the chips OIM chips get designed. Today, weâre changing that.
1
5
40
New blog post: The most secure chip in the world just got hacked. Infineon's SLE78 is one of the bestselling secure elements on the market. It's the chip inside of Yubikeys. And it just got hacked. How did this happen, and how do we stop it from happening again? âŹď¸
2
6
40
the best feeling in the world is realizing thereâs an IEEE paper from the late 90s with a rigorous mathematical analysis of the exact problem you want to solve.
1
1
39
@animalologist either this is an urban myth, we met the same stoner, or this is something that happens surprisingly often.
3
0
38
this is absolutely nuts! they got state of the art efficiency for FPGA CIFAR-10 by training gates, rather than parameters. this also eliminates the need for expensive memory accesses â the gates are the weights!.
Excited to share our NeurIPS 2024 Oral, Convolutional Differentiable Logic Gate Networks, leading to a range of inference efficiency records, including inference in only 4 nanoseconds đď¸. We reduce model sizes by factors of 29x-61x over the SOTA. Paper:
0
0
38
this is a pretty big misunderstanding of the chip design workflow. LLMs wonât offer huge performance improvements for silicon. But they will be able to significantly reduce the engineering overhead to execute on an ASIC project, especially on the verification side.
2
1
36
In the 1950s and 1960s, researchers built computers using a low cost, error-robust architecture invented by John Von Neumann: stochastic computing. But as chips got better and cheaper, it fell out of favor. Now, the computational needs of AI might bring it back. đ§ľ
1
4
35
me hiring for a talented engineer who understands both analog and digital design.
2
0
35
come design chips at normal computing. rent a walk up on the upper west side. hang out on the giant terrace with your cats and a glass of bourbon. i did it you can too anon
3
0
35
âour plan to beat nvidia for transformer inference isâŚâ. stop. no. donât try to be better at matrix multiplies than nvidia. they have 10,000 geniuses all trying to beat you. find a different workload theyâre bad at and focus on that instead.
6
0
33
new blog: cerebras' is trying to IPO!. they have a very cool product and exciting revenue growth -- but 87% of that revenue comes from G42, an UAE-based company with questionable ties to China. why won't any other big cloud players buy their chips?. read more âŹď¸
4
1
31
New blog post!. Interest is growing in secure AI chips, but there's not an efficient way to efficiently implement side-channel secure AI chips without consuming tons of silicon area or power. But there may be a new solution: approximate arithmetic. đ§ľ
2
4
33
y combinator is wrong about LLMs for chip design. on the blog this week, we're analyzing why LLMs won't be replacing silicon design engineers, and won't enable 100x faster silicon. but they do still represent a massive opportunity. đ§ľ.
1
4
32
@thechosenberg there's a very strong correlation between the quality of my relationships and how quickly we became officially a couple. my happiest, longest relationship went from meeting to being a couple in 6 days.
1
0
30
Etched just raised $120M to build chips that are only good at running transformers. But their announcement stretched the truth quite a bit⌠đ§ľ
3
1
30
protagonist ventures suddenly pulled out of a round we were doing, a day or two before closing. not illegal, but i would never take their money again.
We need to make space for exited founders to call out the VCs who wronged them. There are so many repeat offenders that new founders don't know about. But once founders get liquid they act like it's water under the bridge, and the cycle continues.
1
0
30
are you a fan of bayesian inference? because id love to sample your posterior.
7
1
30
yc used to be for outsiders and misfits who thought they could build something awesome through hard work and grit. now itâs for elitists who think theyâre smarter than you because they built a GPT wrapper.
An uncomfortable truth: a small minority of people are dramatically smarter and more effective than average. Most folks have never seen these people operate first-hand. It's both awe-inspiring and humbling. These people aren't always particularly "relatable" or "electable", so.
3
2
28
new blog post!. - i finally swung by @fractaltechnyc & it's great!.- my team at @NormalComputing is hiring, come work with me!.- various thoughts about the best first customers for startups đ¤đ°. âŹď¸.
2
1
28
most undergrad computer architecture classes focus on CPUs and control-flow-based thinking. but the most important skill for modern accelerator design is dataflow-based thinking.
1
1
28
@alexisxrivas @coverbuild the kids living with their parents are by and large not the same kids whose parents can afford to build them a mini house.
1
1
27
unfortunately the startup i work at is going to start winding down. if anybody has opportunities for IC designers with experience designing accelerators for AI and/or cryptography, shoot me a DM!.
8
0
29
why would i take a pill that makes me worse at banjo, lifting, or flirting.
2
0
28
@thechosenberg in the best relationship i ever had, we went from our first date to being officially a couple after 6 days, having gone on 3 dates in which we spent nearly 24 hours together total. honestly i just want that again, it was incredibly easy.
1
0
26
It costs more power to flip a register from 0->1 than it costs to keep it at 0. This lets attackers extract the Hamming weight of secret keys from the chip. Read more in the full paper!.
2
1
27
very excited to get to be a part of this! ARIA is an awesome agency and they're funding tons of cool projects. if you wanna help us build next-generation compute, we're hiring silicon people in the US and the UK!.
Current mechanisms for training AI systems utilise a narrow set of algorithms/hardware building blocks and require significant capital to develop + produce. In Scaling Compute, weâre funding 12 teams of Creators in an effort to reduce the cost of AI hardware by >1000x â (1/2)
2
2
27
my parents are flying to nyc on Christmas Eve and showing up at my small apartment at 8pm for dinner. hereâs my menu:.- whitefish salad with crackers.- roast bone marrow with sumac onions and toasted sourdough.- miso roast pork shoulder with cranberry sauce & salad.- babka.
4
0
26
> viral post about low precision training for LLMs. > look inside. > MNIST.
wrote a paper: it lets you *train* in 1.58b! could use 97% less energy, 90% less weight memory. leads to a new model format which can store a 175B model in ~20mb. also, no backprop!
2
0
26
@RiverTamYDN I mean she was an asshole about it but I totally get not wanting to date this guy. Itâs kinda duplicitous of him, which is a red flag, and she probably cared about the degree because she wants to date within her social class.
0
0
23
thanks for 2k followers!. more hardware startup content coming soon but for now enjoy this vignette of reasons to move to new york
0
1
26
i want to fall in love with a girl in a mosh pit and take her home and have her tell me sheâs a successful corporate lawyer.
2
0
26
when i was finishing college i got a new grad offer to work on the tensor core. iâm glad i didnât take it but sometimes i wonder what could have been.
78% of $NVDA employees are now millionaires. 1 in 2 is worth over $25 million.
0
0
26
kill: VHDL. kill: verilog. kill: systemverilog.
Letâs play fuck/marry/kill with programming languages. I would. Fuck: Haskell (beautiful language but long term hard to build real software). Marry: Rust (beautiful lang, requires some comprise, but overall faithful and safe). Kill: Python (you already know why). What about you?.
4
0
22
@the_aiju @LeetAlpaca3 this is the ideal pcb. you may not like it but this is what peak performance looks like.
2
3
23
Ising machines -- analog solvers for optimization problems -- are a hot area of research. But what if you could use them for more than just optimization? đ§ľ
1
1
22
pat gelsinger got ousted from intel and immediately started ratioâing random analysis on twitter.
@dnystedt speaking about yield as a % isn't appropriate. large die will have lower yield, smaller die - high yield percentage. Anyone using % yield as a metric for semiconductor health without defining die size, doesn't understand semiconductor yield. yields are represented as defect.
3
1
25
@idoccor @albfresco @ctjlewis transistors perform best if theyâre aligned to the crystal lattice of the silicon, which is square! so at least the transitors themselves need to be at right angles. also I think it would be harder to build EDA tools for hexagonal layouts?.
1
0
21
the value of dense compute is underappreciated. dense compute -> more room for on-chip memory -> fewer off-chip memory accesses -> faster, more efficient chips.
5
0
24
every AI chip startup is focused on inference right now, but training reasoning models with RL could very well be a new opportunity for hardware acceleration.
1
0
24
this isnât a chip.
Introducing Oasis: the first playable AI-generated game. We partnered with @DecartAI to build a real-time, interactive world model that runs >10x faster on Sohu. We're open-sourcing the model architecture, weights, and research. Here's how it works (and a demo you can play!):
2
0
24
make chips to accelerate ai to make chips to accelerate ai to make chips to accelerate ai to make chips to accelerate ai to make chips to accelerate ai to make chips to accelerate ai to make chips to accelerate ai to make chips to accelerate ai to make chips to accelerate ai to.
3
1
21
as an ai chip architect youâre faced with an important question:. which long-dead computing paradigm will you attempt to drag from its grave and sculpt into a twisted simulacrum of a novel architecture.
0
1
22
when normal computing chips replace nvidia chips i promise you this:. i am going to make a programming framework thatâs so much worse than cuda just to torture software engineers.
2
1
23