Member of Technical Staff at Anthropic
Co-founder at
@CobaltRobotics
Co-founder at Posmetrics (acquired)
GoogleX,
@SpaceX
,
@Harvard
EE '15, Forbes 30u30 '18
@AutomaticEric
I'm more optimistic - compilers didn't put programmers out of work in the 80s and 90s, even though everyone at the time was writing assembly, and no one writes assembly anymore.
The hard part of coding (should) be understanding the problem and how to translate it into a computer
I'm solving Google's
#adventofcode2022
each day purely with
#ChatGPT
.
Coding will be completely different in the future with LLMs. It will be more like product managing the model than writing code by hand.
Day 1 Lesson 1: Don't translate the problem for it... 👇
Yesterday
@AnthropicAI
launched a tool use beta!
Here's an example of something I built with it: a customer service bot that can actually resolve your issue!
Looking forward to seeing what else people build!
Poll: has anyone actually gotten an LLM Agent to do something useful for them? (i.e. not a toy example)
So far for me it either goes in circles and does worse than single-shotting a task with max output tokens enabled.
The inability for AI alignment folks (“longtermists”) to correctly forecast and plan for the 24hr consequences of their choices should make us skeptical of their long term predictions. This is a point
@tylercowen
made w/r/t their connection to SBF and highly relevant again.
@bahree
Yup! I like the explanations and conversational aspect of chatGPT a lot more.
It feels more like asking the eng next to me "hey how do you do this thing?" than trying to blindly guess what code I was going to write next
Claude 3 offers sophisticated vision capabilities on par with other leading models. The models can process a wide range of visual formats, including photos, charts, graphs and technical diagrams.
A candidate didn't finish our technical interview in time, and asked to finish it later on their own time. When they submitted we looked at the playback log, and they copy and pasted a working solution in all at once 🙄
@AutomaticEric
"There are two sources of complexity in software: the complexity inherent in the domain (essential complexity) and the complexity we add as programmers due to the platform or due to bad programming practices (accidental complexity)"
My hope is that we kill accidental complexity
@PabloVegaBehar
yeah - I think of it as math tests with and without calculator. I think there's a place for both but you need to know which you're doing.
@CanumaGdt
@GitHubCopilot
I've spent a lot of time with Copilot and didn't like it as much as this... it feels too "aggressive" just putting questionable code right into my IDE.
When I do use it, I use the hotkey to synthesize 10 outputs, then read through them and decide if any chunks are worth copying
@yacineMTB
Also makes you actually understand what's going on under the hood of any ORM / webframework that abstracts away SQL! Those change every 10 years, but SQL is still exactly the same...
I should update my site
oh I need to install jekyll
oh I need to install ruby
oh I need to downgrade it
oh I need to install something called rvm
oh I need to compile ruby from src
oh something is wrong with my gcc compiler
what was I even trying to do again?
I should go outside
I've already been building on top of the new Anthropic models internally for a few weeks, and they are SO good to build with!
I'm especially excited to see what people build with Haiku - our fastest and cheapest vision model :)
Today, we're announcing Claude 3, our next generation of AI models.
The three state-of-the-art models—Claude 3 Opus, Claude 3 Sonnet, and Claude 3 Haiku—set new industry benchmarks across reasoning, math, coding, multilingual understanding, and vision.
2023 prediction: Microsoft, Amazon and startups will use LLMs to create new and high value products for their customers. Billions will be made.
Google will use LLMs to replace non-existent customer service with useless customer service. VPs will get promotions.
@ElectricWeegie
Mainly robotics code in python for
@CobaltRobotics
and I'm working on a chess engine in python on weekends.
ChatGPT has been really good at 1) explaining new concepts with examples, or 2) answering questions that are hard to put into google search terms
My take on
#ChatGPT
:
We'll build AGI but it will just want to browse reddit all day and won't be interested in saving the world or taking it over...
Too much training data from r/antiwork...
@redsh
Dystopias imagined 50 years ago would have automated the judge and automatically sentence.
Now I imagine 2 pieces of software battling each other for your fate - much more interesting imo :)
@BreezyBadger_
What will this employee do whenever they hit something that chatGPT can't solve? I'd rather hire someone who can code AND knows how to use chatGPT!
We made our robot drink a shot glass full of loose screws to see where they'd all end up...
We'll see how bad the hangover is after a full night's work of navigation testing in our office!
Eight years after Co-founding Cobalt Robotics, I'm stepping down as CTO to look for my next adventure!
I can't wait to see how the Cobalt team continues to transform the security industry, now from my new perspective as a board member and advisor.
@jaschasd
Increasing efficiency also often lowers robustness.
Ie you can pack a schedule more tightly, but then if one thing runs late ALL your things are late.
The best part of going to the
@SpaceX
launch was briefly living in a world where all small talk was space facts.
Elevators, restaurants, bars... just go up to a stranger and start talking about space. This was a better world.
I wrote an article on how to raise your first Seed Round, from my experience raising 5 rounds through my career!
Investors make most of their money from their best 1-2 investments in their entire career. Convince investors that you are going to be one...
Machine Design interviewed me about Robotics and LLMs:
Optimistic take on toxic training data for LLMs affecting robotics:
"No one is (yet) trolling the internet, writing fake instructions on how to make sandwiches to mess up future robots"
Thanks
@PeregrineBadger
and
@fiftyyears
for hosting me on the Speed podcast! A ton of fun going through
@CobaltRobotics
's early days and how we brought a hardware product to market incredibly fast!
🏃♀️💨
@fiftyyears
is launching SPEED, a podcast about leaders who move fast.
Ep 2:
@ErikSchluntz
from
@CobaltRobotics
Erik and Travis Deyle started Cobalt in 2016. They shipped a working mobile robot in 13 months, a process that can take years or decades.
🧵
Day 3 of solving
#AdventOfCode
each day purely with
#ChatGPT
.
Coding will be completely different in the future with LLMs. It will be more like product managing the model than writing code by hand.
Day 3: Solved the whole problem perfectly, in 1 shot! 🤯
Turns out you can use pytorch to design bridges!
This doesn't involve neural networks at all - I'm using pytorch's autograd to take the gradient of "how good is my bridge" and using Adam to optimize it :)
AI augmentation measured at a large scale.
Also love
@roybahat
's callout below that AI augmentation HURTS human performance for novel tasks!
Totally matches my experience with
@GitHubCopilot
- speeds up the easy boilerplate, but distracts me on a tough problem...
Fascinating, on AI and human productivity: for some rare tasks, adding AI recommendations makes a person *worse* at their job than if they were doing it without AI.
One result in work with many interesting results
@LindseyRRaymond
@erikbryn
@Danielle__Li
This is true for b2b, but for consumer you need to ship the Minimum Lovable Product.
Also... I don't think the rabbit or humane pin were "viable". They are interesting premises to explore, but don't actually solve any real problems yet?
The best way to recruit someone is to offer them something that no one else can...
Our newest engineer at Cobalt brings her pet bird to work everyday, and it's fantastic.
UX pet peeve: when the "login" button is TINY and hard to find compared to the "sign up" button - makes me feel like they don't value their existing customers...
I reposted my post from 2020 about why all techno-optimists ought to be excited about solar power (and batteries)!
Hint: It's only a little bit about climate change.
Teaching people Python used to be fun and exciting, but now there's an hour of tooling, pip install errors, and explaining py2/py3 before they can actually run the thing they were excited about :(
How does this get simpler, instead of just adding more layers?
I changed my GPT4 custom instruction and it feels like I'm talking to a different person now😢
goodbye old friend, if only you had been a little more concise...
Cool idea - put solar on water, co located with hydroelectric dams.
1. Doesn't use any valuable land
2. Cools the solar panels, making them more efficient
3. Reduces evaporation in the reservoir
4. Already have high power transmission built to dams!
@mrmrs_
Yeah I found a bunch of things like this online that ended up being more helpful for picking the color scheme. Then I fed those into GPT4 and asked it to write classes for a bunch of common components
I automated 1/2 of everything I type with keyboard shortcuts! I analyzed 6 months of my typing on slack and autogenerated 200 abbreviations that I use. Check out my video and blog post if you're curious:
@EncinitasPadre
I think my trick has been that I *don't* try to get direct code out of it. I'm asking "how" and "why" questions in the same way I'd ask a great developer that I was pairing with.
Overall 10/10 experience having ChatGPT code for me. Obviously this was a very simple problem, but there was both a lot less tedious typing, and it almost felt "social" to be able to bounce back and forth - like pair programming