Rohan Relan (@hanrelan) profile picture

Building ScribbleVet Formerly: Founder @ Agawi (acq by Google), Pluto Media (acq), Miru (YC W18). Tech Lead @ Google App Streaming

San Francisco
Joined May 2009
Photo by Rohan Relan (@hanrelan)Photo by Rohan Relan (@hanrelan)Photo by Rohan Relan (@hanrelan)Photo by Rohan Relan (@hanrelan)Photo by Rohan Relan (@hanrelan)Photo by Rohan Relan (@hanrelan)Photo by Rohan Relan (@hanrelan)Photo by Rohan Relan (@hanrelan)Photo by Rohan Relan (@hanrelan)Photo by Rohan Relan (@hanrelan)
Regularly updated meta thread tracking my recent AI explorations & experiments 👇
1
2
16
2,898
David's protein milkshake has 2 grams of fat in it and tastes like a milkshake. The fat is one ingredient, EPG, and your body mostly doesn't absorb it. The label sorts itself into four systems, which is unusual for a consumer product. Protein system: ultrafiltered skim milk, water, lactase enzyme. Fat system: modified plant fat, EPG. Flavor system: Dutch process cocoa, natural and artificial flavor, salt, sucralose, acesulfame potassium. Structure system: dipotassium phosphate, gum arabic, cellulose gum, gellan gum. EPG is the interesting part. It starts as rapeseed oil. They split the oil into glycerol and fatty acids, insert propylene glycol units into the glycerol, and put it back together. Those inserted units sit right where lipase needs to grab, so the enzyme mostly can't break it down. It coats your mouth like fat and carries flavor like fat, then leaves without giving up its calories. 10 grams of regular fat is 90 calories. 10 grams of EPG is about 7. If you're old enough to remember olestra, that's the right comparison and it's the reason this one is built differently. Olestra was a liquid oil at body temperature, and that's where the side effects came from. EPG melts at 102 degrees, above body temperature, so it stays solid on the way through. The company describes the consistency as soft candle wax. The structure system is four ingredients. Gellan gum comes from a bacterial fermentation and forms a gel so thin you can't feel it, but strong enough to hold cocoa particles in suspension. It's what keeps the cocoa from sinking to the bottom during months on a shelf. Gum arabic is dried sap from the acacia tree. It's an emulsifier, so it keeps the fat and the water from separating. Cellulose gum is plant fiber that's been modified to dissolve, and it adds body so the shake feels thick instead of watery. Dipotassium phosphate is a buffer. Milk protein wants to clump when it's heated to make a drink shelf stable, and the phosphate holds the pH where the protein stays dissolved. The sweeteners are a pair. Sucralose is slow to hit and lingers. Acesulfame potassium hits fast and has a bitter finish. Blended, each one covers the other and the result tastes close enough to sugar. The Dutch process cocoa is cocoa treated with alkali, which darkens it, softens the bitterness, and makes it disperse in liquid instead of clumping the way natural cocoa does. EPG is GRAS, not an approved food additive. The company assembles the safety file, FDA reviews it and issues a letter saying it has no questions. There are three of those letters on file. The molecule was developed in the 80s and didn't reach a product until around 2020. The longest human study in the file ran eight weeks. So the safety data is real, but the long term impact is being written right now by the people drinking it.
112
102
69
2,988
303,835
Rohan Relan retweeted
In medieval times, within the arms race of ever more demonic torture devices, some sadistic genius came up with the idea of the Little Ease. This was a prison cell built so small in every dimension that a grown man could not stand upright in it nor lie down at full length nor properly sit. The pain is relentless and without relief and inflicted by one's own body. Prisoners were known to go insane within a few days. A stay at the Little Ease was considered even more cruel than the rack, the thumbscrew, and the other ghoulish machinery of the Tower of London. A breeding pig will spend her whole life in a version of that box. These are social, roaming creatures (more intelligent than dogs) who will never leave this corset of steel. They have been selectively bred to be bigger than their frames can support. Yet we put them in cells so confined that they cannot comfortably sit, and their attempts to do so (for example, by sneaking their limbs into adjacent stalls) reliably lead to fractures and sprains. They cannot sweat, yet have nothing to roll around in to cool themselves off. Except their own manure, which (contrary to the common misconception) they are so averse to (thanks to their strong sense of smell) that new sows will often suffer from constipation to avoid soiling the space from which they eat and sleep. Here is how the writer Matthew Scully described what saw at one of Smithfield’s “gestation barn”: > “Sores, tumors, ulcers, pus pockets, lesions, cysts, bruises, torn ears, swollen legs everywhere. Roaring, groaning, tail biting, fighting, and other “Vices,” as they’re called in the industry. Frenzied chewing on bars and chains, stereotypical “vacuum” chewing on nothing at all, stereotypical rooting and nest building with imaginary straw. And “social defeat,” lots of it, in every third or fourth stall some completely broken being you know is alive only because she blinks and stares up at you … creatures beyond the power of pity to help or indifference to make more miserable, dead to the world except as heaps of flesh into which the [insemination] rod may be stuck once more and more flesh reproduced.” — The Save Our Bacon Act is trying to unroll the few state protections we have against this barbaric cruelty - for example California’s Prop 12 - which banned the sale of pork from pigs kept in gestation crates. It’s incredibly important we don’t end up with this sort of federal preemption. SOB will not only kill the most important animal welfare related laws in the US of the past decade, but more importantly, it will also restrict ALL future legislative progress (aka how the animal welfare movement has gotten its biggest wins). The Senate is currently deciding whether to add the SOB Act to the Farm Bill. With relatively little money now, we can discourage the most pivotal senators in the Ag committee from backing this amendment. Defeating this bill is even more important given the amount of philanthropic funding I expect to come online in the next year or two. It will plausibly be over 10x more expensive to repeal SOB than to prevent it from passing in the first place. All that money that could be spent transforming our society's relationship to mass animal suffering will instead have to be spent just getting us back to where we are right now. That's why money spent now fighting this bill (and I mean right NOW) is so effective. If you’re in a position to donate six figures, please DM me.
99
773
72
4,512
542,392
Rohan Relan retweeted
Hell isn’t a theoretical concept, we have willingly created it for billions of beings not that dissimilar to ourselves
In 2023, Chinese researchers set out to create a 'pig depression model' — a way to reliably induce depression in pigs for medical research. The researchers (Yang et al, 2023) confined pigs so tightly they couldn't move and left them in a dark environment without food or water for 24 hours. It worked: the pigs exhibited anhedonia — an inability to experience pleasure — along with anxiety and reduced appetite. The researchers concluded they'd found a promising way to induce depression in pigs. Their inspiration? Factory farms. They tried out the total restraint after noting that sows, confined in gestation crates and kept chronically hunger, "have been found to be in a depression-like state." The main difference? After 24 hours, the researchers let the study pigs out. On factory farms they never do. Photo: Balvik C. / We Animals
42
825
6
8,313
281,573
Rohan Relan retweeted
Hidden on page 744 of the farm bill the House Agriculture Committee passed Thursday is a provision that would condemn millions of pigs to a lifetime in gestation crates. Rebranded the 'Save Our Bacon Act,' it's a pork-industry play to wipe out every state law banning the sale of pork from crated pigs — laws the conservative Supreme Court upheld in 2023. Over 85% of Democrats and Republicans oppose these crates. Voters have backed ballot measures to ban them in state after state. The pork industry knows it can't win a straight vote on this. So it's burying the provision in an 800-page bill and hoping no one notices. Contact your senators and representative today and tell them: oppose the farm bill unless the Save Our Bacon Act is stripped out. You can reach them at senate.gov and house.gov — it takes two minutes and it matters.
1,059
11,819
1,134
25,859
5,289,879
Rohan Relan retweeted
Today, we're announcing our $33M Series A led by Dawn Capital. Oneleet is ending compliance theater by making companies genuinely secure. Compliance follows automatically. Here's our story. 🧵
42
20
8
311
111,682
Bad news (1/2): video taken down by ICML ([email protected]) for copyright. While I can't agree (the consent I signed allows me to publish elsewhere) - I will respect it to save time for more important things. To bad I delayed many things and spent 20+ hrs preparing the video.
Incredibly honored and humbled by the overwhelming response to my tutorial, and thank you everyone who attended in person. Truly heartwarming to hear how much you enjoyed it. Many have been asking for a recording, and I prepared one with my own subtitles youtu.be/yBL7J0kgldU
35
31
15
471
238,069
"finishReason": "RECITATION" How are folks using Gemini and avoiding this in production? Would love some pointers here (maybe @OfficialLoganK has some tips?)
1
1
152
📽️ New 4 hour (lol) video lecture on YouTube: "Let’s reproduce GPT-2 (124M)" youtu.be/l8pRSuU81PU The video ended up so long because it is... comprehensive: we start with empty file and end up with a GPT-2 (124M) model: - first we build the GPT-2 network - then we optimize it to train very fast - then we set up the training run optimization and hyperparameters by referencing GPT-2 and GPT-3 papers - then we bring up model evaluation, and - then cross our fingers and go to sleep. In the morning we look through the results and enjoy amusing model generations. Our "overnight" run even gets very close to the GPT-3 (124M) model. This video builds on the Zero To Hero series and at times references previous videos. You could also see this video as building my nanoGPT repo, which by the end is about 90% similar. Github. The associated GitHub repo contains the full commit history so you can step through all of the code changes in the video, step by step. github.com/karpathy/build-na… Chapters. On a high level Section 1 is building up the network, a lot of this might be review. Section 2 is making the training fast. Section 3 is setting up the run. Section 4 is the results. In more detail: 00:00:00 intro: Let’s reproduce GPT-2 (124M) 00:03:39 exploring the GPT-2 (124M) OpenAI checkpoint 00:13:47 SECTION 1: implementing the GPT-2 nn.Module 00:28:08 loading the huggingface/GPT-2 parameters 00:31:00 implementing the forward pass to get logits 00:33:31 sampling init, prefix tokens, tokenization 00:37:02 sampling loop 00:41:47 sample, auto-detect the device 00:45:50 let’s train: data batches (B,T) → logits (B,T,C) 00:52:53 cross entropy loss 00:56:42 optimization loop: overfit a single batch 01:02:00 data loader lite 01:06:14 parameter sharing wte and lm_head 01:13:47 model initialization: std 0.02, residual init 01:22:18 SECTION 2: Let’s make it fast. GPUs, mixed precision, 1000ms 01:28:14 Tensor Cores, timing the code, TF32 precision, 333ms 01:39:38 float16, gradient scalers, bfloat16, 300ms 01:48:15 torch.compile, Python overhead, kernel fusion, 130ms 02:00:18 flash attention, 96ms 02:06:54 nice/ugly numbers. vocab size 50257 → 50304, 93ms 02:14:55 SECTION 3: hyperpamaters, AdamW, gradient clipping 02:21:06 learning rate scheduler: warmup + cosine decay 02:26:21 batch size schedule, weight decay, FusedAdamW, 90ms 02:34:09 gradient accumulation 02:46:52 distributed data parallel (DDP) 03:10:21 datasets used in GPT-2, GPT-3, FineWeb (EDU) 03:23:10 validation data split, validation loss, sampling revive 03:28:23 evaluation: HellaSwag, starting the run 03:43:05 SECTION 4: results in the morning! GPT-2, GPT-3 repro 03:56:21 shoutout to llm.c, equivalent but faster code in raw C/CUDA 03:59:39 summary, phew, build-nanogpt github repo
391
2,147
414
15,213
1,529,059
I used to read every support email, but that's not possible anymore. So I wrote a script to give myself and the rest of the team a sense of all the feedback, bugs, and requests coming in. It uses GPT-4 to summarize the last day's worth of our @FrontHQ inbox. Link below 👇
1
3
246
Feel free to use and tweak this! gist.github.com/hanrelan/9b4… Will continue to share more AI-powered internal ops tools as we build them. AI provides so much leverage here.
97
GPT-4 is a perfect match for @nixos_org . It understands the language and the error messages, and since it's Nix you can trivially undo any mistakes. Used it to create a shell.nix for @elixirlang 1.16 in just a few minutes. Someone please make a GPT specifically for Nix!
2
233
Rohan Relan retweeted
🏠Announcing our $115M pre-seed & early-stage Fund III for @Cal founders🍄🧵 nitter.cf/TheHouseVC/status/1717…
Announcing our new $115M pre-seed and early-stage Fund III—dedicated to startups from the remarkable 600,000-person-strong @Cal network, based in Berkeley and worldwide.🧵 tcrn.ch/3FxHcT0
38
35
29
272
109,510
Rohan Relan retweeted
Plugin #2 - HttpProxy Say goodbye to Postman + curl. Just ask ChatGPT to request whatever you want!
8
20
3
171
30,330
Rohan Relan retweeted
Plugins for processing a video clip, no ffmpeg wizardry required. Actual use-case from today's launch.
327
1,605
707
11,312
4,779,983
Wow! Could easily see chatgpt becoming the new front page to the internet. Openai's really going after Google's jugular. Also what a distribution channel for new startups! @carrd / @webflow where's my build a website plugin?!
We’ve added initial support for ChatGPT plugins — a protocol for developers to build tools for ChatGPT, with safety as a core design principle. Deploying iteratively (starting with a small number of users & developers) to learn from contact with reality: openai.com/blog/chatgpt-plug…
1
1
8
528
Well, that was fast… I just helped create the first jailbreak for ChatGPT-4 that gets around the content filters every time credit to @vaibhavk97 for the idea, I just generalized it to make it work on ChatGPT here's GPT-4 writing instructions on how to hack someone's computer
106
656
186
4,275
1,771,669
In the short term, do some programming languages become more interesting because #GPT4 is so good at writing code? If the cost of writing fairly generic code (sdks etc) is so close 0, I'd expect languages that offer more than just a large existing ecosystem would benefit
2
2
325
In the medium/long term the answer will probably be "English" and other natural languages as directly writing code becomes obsolete.
1
87
Maybe assembly will make a comeback for maximizing performance
84