Alex Sullivan (@alexsullivan444) profile picture

Mobile app development. Lots of Android. I write some books. I've been known to play some video games. Sometimes I read. That sort of thing.

Boston
Joined August 2015
Photo by Alex Sullivan (@alexsullivan444)Photo by Alex Sullivan (@alexsullivan444)Photo by Alex Sullivan (@alexsullivan444)Photo by Alex Sullivan (@alexsullivan444)Photo by Alex Sullivan (@alexsullivan444)Photo by Alex Sullivan (@alexsullivan444)Photo by Alex Sullivan (@alexsullivan444)Photo by Alex Sullivan (@alexsullivan444)Photo by Alex Sullivan (@alexsullivan444)Photo by Alex Sullivan (@alexsullivan444)
Alex Sullivan retweeted
📣 for the first time in almost 2 years I have an opening for new clients (in May)! Do you (or someone you know) need help with Elixir/Phoenix? How I help teams: - ⚡move faster building high-quality, well-tested apps - ✨help you ship software that matters to you and your customers - 🚵‍♂️ 🚵‍♀️ improve your team even as we improve the product (remember, your team is a product) - 🍐💻 mentor developers through workshops and pair-programming DMs are open or email me at german(at)germanvelasco.com For more on how I work, you can see germanvelasco.com/#how-i-wor… #MyElixirStatus
1
14
1
37
6,559
We have a brand new website! Stay tuned for some new features. lilacml.com
3
1
21
6,173
We just made it much easier to load and host any HuggingFace dataset in Lilac, with no code! Just provide the Lilac Deployer UI: ✅ HuggingFace dataset name ✅ Your HF access token We will spin up a HuggingFace space with Lilac on that dataset. huggingface.co/spaces/lilaca…
4
18
2,329
Alex Sullivan retweeted
Lilac: 3 minute introduction! This quick video will show you how to use @lilac_ai to *curate* a dataset, removing unwanted responses with "As a language model I cannot..." ✅ Find all the responses with language model self-reference using the built-in Lilac concept ✅ Label them as "drop" ✅ Export the curated dataset, dropping unwanted responses By combining concepts, labeling, and exporting with exclusion, we can now steer the language model to avoid these types of outputs!
1
6
23
5,827
Alex Sullivan retweeted
Just watched @germsvel's great talk on domain driven design. I don't know Elixir, but this talk still had so much to say about how we structure our code. German is one of the most thoughtful developers I know and I always love hearing how he thinks. youtube.com/watch?v=JNWPsaO4…
1
4
16
2,030
Alex Sullivan retweeted
👋 I’m available for a new engagement starting Nov. If you need help building high quality Elixir/Phoenix apps, mentoring developers, TDD training, or building your next product, let’s chat! DMs are open, and retweets would be lovely! germanvelasco.com/ #MyElixirStatus
40
1
64
9,785
Alex Sullivan retweeted
Organized, manual inspection is critical for data quality. At Google we found the silliest *bugs* from this process, ones that are very hard to fully automate beforehand: - Rater instruction confusion: confusing “human race” with “running race” - Question / Answer pairs swapped accidentally - Labels leaking in images Many other examples, but these are easy to fix once you’ve seen them, and often they dramatically boost model performance after being fixed.
One pattern I noticed is that great AI researchers are willing to manually inspect lots of data. And more than that, they build infrastructure that allows them to manually inspect data quickly. Though not glamorous, manually examining data gives valuable intuitions about the problem. The canonical example here is Andrej Karpathy doing the ImageNet 2000-way classification task himself. And in the era of large language models, manually examining data is probably even more insightful since completions are hard to evaluate via benchmarks. In this spirit, I recently did a few days of pair programming with @hwchung27 where we were starting on a new problem. Instead of trying to replicate baselines and design new methods, we ran some evaluations and manually inspected them to gain insights. We first paid about one day of overhead getting all the relevant information in a single UI so we could examine the data without having to click through multiple web pages. The second day, we spent an afternoon reading examples together and taking notes on the patterns that we noticed in the examples. ChatGPT generates long text, and we actually read the whole thing carefully, even if one example took 20 minutes to understand. I think we both gained a deeper understanding of the problem that we could not have gotten from reading research papers. (In 2018, for example, I helped pathologists label a lot of data to train a lung cancer classifier. After having manually labeled 200+ images (with pathologist correction), I’d probably gained a pathologist-level understanding at that one particular lung cancer classification task :))
3
2
22
3,819
Alex Sullivan retweeted
🌸 Excited to announce this 1st class integration of @lilac_ai with Langsmith + @langchain We released a source with 0.1.0 that ingests Langsmith datasets so you can: 🔎 explore user interactions ✨ annotate with concepts & signals ✏️ label & curate for LLM fine tuning
🛠️When building with LLMs, nothing replaces a strong data flywheel.🔄 🌸Lilac helps you see 👀 and quantify 🔢 your unstructured datasets. Now it integrates with LangSmith, so you can go from prototype to fine-tuned application in a flash ⚡ 👉 blog.langchain.dev/fine-tune…
1
6
1
31
5,543
Alex Sullivan retweeted
Replying to @WillManidis
We’re starting to do that with Lilac! Just a few indexed now, but one day all of HF 🚀
We are excited to announce Lilac (@lilac_ai), a new Open Source product to analyze, structure and clean data with AI. Our mission is to make unstructured data visible, quantifiable, and malleable. 🌸 Announcement blog: lilacml.com/blog/introducing…
1
8
854
Trillion dollar idea: AI feature where you can long click a link and get a summary of the contents without going into the link. Click bait solved.
4
223
Might just be that I'm not able to get past personal habbits, but I've had the generative AI thing enabled in Google for a while and have gotten exactly zero value from it. Mostly it feels like an annoying pop-up. Is anyone getting anything from this?
1
134
Alex Sullivan retweeted
Replying to @Altimor
We’re working towards this
We are excited to announce Lilac (@lilac_ai), a new Open Source product to analyze, structure and clean data with AI. Our mission is to make unstructured data visible, quantifiable, and malleable. 🌸 Announcement blog: lilacml.com/blog/introducing…
1
6
629
Alex Sullivan retweeted
We are excited to announce Lilac (@lilac_ai), a new Open Source product to analyze, structure and clean data with AI. Our mission is to make unstructured data visible, quantifiable, and malleable. 🌸 Announcement blog: lilacml.com/blog/introducing…
25
125
22
656
233,662
Every time I get back into iOS development I'm blown away, in a bad way, by the state of developer tooling. Spicy take - iOS development felt _a lot_ better before Swift came along. Now everything feels needlessly complicated and painful.
2
1
162
Alex Sullivan retweeted
If you work with text data that gets fed into an ML model and spend time cleaning data, we'd love to chat and give you a peek at our new product. DM me! @dsmilkov and I are happy to share advice and insights we've learned from working with ML data at Google.
2
7
2
30
8,529
Alex Sullivan retweeted
🛍 Our Spring Sale is on — 40% off books, and get our Kodeco Personal 17,520 Hour Plan (or call it our Two-Year Plan, same diff) for just $599 for a full two years of access to the best and highest-quality mobile dev resources anywhere. #iosdev #androiddev hubs.ly/Q01MRVcy0
3
20
2
18
8,020
The US is the 6th-most expensive country in the world to build rapid transit infrastructure, & yet the quality & quantity of transit projects is considerably inferior. 🧵 So, why does it cost so much to build transit in the US?
7
48
5
210
34,397
Excited for when I'll be able to manually type out a filter for social media and have some LLM decide whether the content matches my filter or not. Will make it super easy to filter out stuff i don't want to see. Looking at you tv clips in TikTok. Your days are numbered.
2
481
Statements like these show that it's time for a new board. One that has a vision for the system, wants rider input, and asks questions and shows real engagement with the work.
1
13
3
119
17,811
Alex Sullivan retweeted
Despite how obvious this is, somehow it’s still hard to convince people to rigorously look at their data. With Know Your Data, we pretty much *always* find something surprising when studying datasets. And typically they’re just stupid data bugs, often with free model perf wins.
Manual inspection of data has probably the highest value-to-prestige ratio of any activity in machine learning.
2
7
43
9,382