🗜ML Engineer
🛠️ Building Multi-Modal Models
@AbideAI
🖋 Writing
#LLMOps
- Managing LLMs in Production book (O'Reilly)
🐦 Talk to me about LLMs, MLSys & SecOps
1. Survey of Large Language Models - it summarizes the available resources for developing LLMs choronologically and discusses the open challenges for futher development.
This has been such an excellent year for software system design in ML. So, I compiled a list of some of my favorite papers 📜in MLOps.
Here are some of my favorite ones till date⤵️
The Mathematical Engineering of Deep Learning is easily one of the best books in ML I have read in a very long time. Ch.7 on Sequential Models and ch.8 on Generative Models is what you're looking for. Can't wait to get hard copy when it comes out. 10/10
@poppacalypse
Not every tech person, haha!
I instead wanna live on a mountain overlooking a waterfall and a big lake, with no humans in sight but with excellent internet, my own helipad and all groceries delivered by a self-driving car, approx within an hour drive from a big city just in case
In 2023, Data Engineering is the fastest-growing field in machine learning. A lot of challenges that we have previously explored in database design are now being adapted to Machine Learning.
Here are some of my fav papers on the topic: 👇🧵
Pleased to announce that I am authoring an O'Reilly book on LLMOps. Thanks to Nicole
@NaB79Vrisnik
for making it happen.
The book will cover local as well as API-based models, LLMOps pipeline, infra/tooling as well as threat modeling, security & privacy for your LLM apps.
The top one would be Machine Learning: The High-Interest Credit Card of Technical Debt by D. Sculley et al
We invited him as a guest on our
@mlopscommunity
podcast (Spotify/iTunes) Episode
#32
- def worth listening to!
If there's one I would def read it would be Machine Learning Operations (MLOps): Overview, Definition, and Architecture by Dominik, Niklas and Sebastian.
It highlights necessary principles, components, and associated architecture and workflows in MLOps
I am planning on teaching Section 1 of my ML in Production course on 11th March for a limited test audience for free either via Twitch Live or Patreon Live or YouTube Live (TBD). Haven't figured out those details. Anyone interested? Interactive session, not a lecture.
@nikillinit
There's a big difference between being alone v/s lonely.
You can be lonely amidst big crowds. It's a belonging-ness problem, not a going-out problem.
I can be alone for days and never feel lonely. Date the wrong person and suddenly be lonely, despite being in ppl all day long
@jesss_codes
You can't. We have far too many hobbies and side projects to have time for real relationships except for friends who are part bug fixers and part collaborators on our hobby projects. But if you don't require a lot of time, we are pretty nice people.
A recent one is Operationalizing Machine Learning: An Interview Study by Shreya et al which interviews 18 MLOps practitioners and discusses common practices across different stages of an ML project from experimentation ->deployment-> monitoring.
I'll be speaking at Fine-Tune & Operate Your Generative AI Models with Andreas Welsch this July 6th. We will talk about:
- How to tailor LLMs to your industry or domain
- How you can improve the results and fine-tune an LLM
and more
RSVP 👉
While a guide for academia, but generally applicable best practices for all Data Scientists and ML Engineers is How to avoid machine learning pitfalls: a guide for academic researchers by Michael A. Lones
@shimon8282
Not a professor but I hate giving career advice and mentoring people in general. No one ever has a linear path in life and there are so many confounders that need to be addressed to give unbiased advice. It feels like a disservice to just proliferate our biases to kids.
1. Data Engineering for Everyone by Vijay Reddi, Pete Warden et. al. posit that open-source data sets are the rocket fuel for research and innovation and explore how to accelerate data-set creation via automatic data set generation tools.
2023 has been a crazy year in AI research esp forLLMs and GenAI, so I decided to compile a list of my fav most significant papers that are worth a read🧵
⤵️
As I am almost wrapping up this book Engineering MLOps, I can comfortably say that this is one of the most comprehensive resources on ML Model Deployment in production and esp automating ML pipelines.
If Moore was alive today, he would lose his mind over the computational glass-ceilings we are breaking every 3 months right now.
What a great time to be alive!
@bhutanisanyam1
That's the responses only until you turn 26.. Once you're 26, the response is, "that's great but all that achievement is worth nothing if you don't have a family. Xyz got married and have two kids now. Is anyone even interested in dating you? What do they do?" 🤣
Remember the Chinchilla model paper? It proposed that for compute-optimal training, the model size and the number of training tokens should be scaled equally.
Google derived its own optimal compute-optimal ratios. See how they differ.
Buying an executive chair for my desk so it can support my increasingly longer hours at my desk as my MLOps course finally gets approved and I start working on my dockerfiles and cloud files. To give you an idea, here are some snippets from the ToC
I am thinking of creating an
#LLMOps
podcast.
There's a lot going on in the LLM world. Far too much for one person to keep tabs on. Key idea - unscripted, unedited weekly show with hot takes on all new announcements and releases.
Looking for a co-host.
Of course, the list won't be complete without a discussion of Jupyter NBs. But what would be the performance difference b/w notebooks vs scripts and the pros and cons of each?
📜A Large-Scale Comparison of Python Code in Jupyter Notebooks and Scripts
Most of the ML deployment challenges in enterprise boil down to conventional software engineering challenges multiplied by dynamic data in an environment where you replace rule-based algos by pattern-based algos.
Today
@6am
- Discussing my upcoming LLMOps report with my 🤩editor Sarah. We have spent about 4 months making sure we can sum up the entire topic in 30 pages as a preface to my book. One more final round of edits this week and it will finally be ready for the world to see 😍
Don't ask me why I hate LinkedIn..
It's this kinda stuff going viral and getting reposts in there 🤦♀️
Who'll tell all these LinkedIn influencers that these aren't "AI skills" but "products/tools with some AI integration"?
@tunguz
I would think they're now instead on or hugging face - i haven't even see kaggle's posts in a while or emails .. idk what their marketing or DevRel teams are upto lately?
Next is a paper that talks about how to address the eng challenges associated with distributed training if u don't have the necessary infrastructure to match the big corps with infinite compute and a million hyperparameters.
Training Transformers Together
⁉ What is it important work in the right direction?
LLMs memorize facts and knowledge contained in the training corpus. Infact, LLMs are not able to recall facts and often experience hallucinations by generating statements that are factually incorrect.
🙏 LLMs+KGs address the above issues. Knowledge graphs (KGs), storing enormous facts in the way of triples, i.e., (head entity, relation, tail entity), are a strstructured and decisive manner of knowledge representation.
2. The Dataflow Model: A Practical Approach to Balancing Correctness, Latency, and Cost in Massive-Scale, Unbounded, Out-of-Order Data Processing by Tyler Akidau et al talks about modern data processing approaches and model semantics for one.
📕The Deep Bootstrap Framework: Good Online Learners are Good Offline Generalizers: The core idea is to couple the Real World, where optimizers take stochastic gradient steps on the empirical loss, to an Ideal World.
Most people think being smart is something you're born with.
I consider myself decently smart but 99% of it comes down to that I just work a lot and read even more and a lot of what I read is always subconsciously building neural pathways connecting existing to new information.
What to do about LLM FOMO?
Learn to be a tester.
When done with testing, get back to the fundamentals-
1. Customer acquisition
2. Database management
3. Putting models in production
4. Personalization
5. Making revenue
Don't get attached to tools, use tools to ur own end!
I've talked to tons of ML Engineers and investors in the space. Everyone agrees that LLMOps isn't just MLOps. Sure, there are some key components that are similar between both but the difference is nothing short of the gap between statistical models vs DL.
I think this is mostly right.
- LLMs created a whole new layer of abstraction and profession.
- I've so far called this role "Prompt Engineer" but agree it is misleading. It's not just prompting alone, there's a lot of glue code/infra around it. Maybe "AI Engineer" is ~usable,…
Last, how can all the progress in machine learning guide the future of chip design? The paper provides an interesting outlook into hardware for software folks.
The Deep Learning Revolution and Its Implications for Computer Architecture and Chip Design
We have all been there. What advice would you give them?
I'll say - The field has changed so much since I got into ML back in 2016/17, my personal roadmap may be irrelevant at best.
What to do now? Do project-based learning. Pick a tutorial from medium. Implement. Improvise.
Always loved the concept of anti-library and have accumulated so many books over the years without any guilt/intention of not reading/finishing them all.
Its just so satisfying to know what else exists out there on my shelf if ever left wanting more.
To do so, they have used several prompting methods.
But overall the evals are done across
1. Sentiment analysis
2. Toxicity detection
3. Text classification
4. Reasoning
5. Coding
6. Language - gen & translation
7. Question-answering
KGs are crucial for various applications as they offer accurate explicit knowledge. Besides, they are renowned for their symbolic reasoning ability, which generates interpretable results.
29 is such a weird age, I can't wait to turn thirty next year just so I can write that tweet-thread of 30 pieces of wisdom 🤣😜
Coz no sensible mature person will tell you the 29 secrets to success, its like you lose all credibility the moment you say "here are my 29 tips"🤦♀️
Books like these included in the core curriculum at universities are essential to help kids really fall in love with technology and become curious about the world around them, more than theoretical books combined with the top 10 highest paying future job prospects.
What my parents think my skillsets are as a tech person?
* I know all phone and app settings by heart and can undo their device to the last settings.
* I can fix routers.
* Operate the surveillance cameras
* Run websites that are down
* Debugging hardware failures
@YakimchukLyuba
But are u sure the data collection for this survey is truly balanced and not biased? I wouldn't jump to that conclusion that quickly based on just a cute chart. Such surveys are often biased with bad practices that can create noise when it comes to data collection & analysis.
But what about Production Infrastructure? How to cater to data stalls in the pre-processing pipeline?
Understanding Data Storage and Ingestion for Large-Scale Deep Recommendation Model Training
We see a new evaluation framework for LLMs daily, but why do evaluations matter?
In this new blogpost,
@ElectricWeegie
and I take a stab at explaining the challenges of evaluating Large Language Models in Production.
While I am no influencer to say **oh yay, 5K followers** but here's my celebration announcement if I were -
You're more than welcome to join me tomorrow for the session I'm teaching to a very limited private audience of ML folks..
Pleasantly suprised by all the work from Apple Research first on Flash Memory () and now on their multimodal LLM Ferret (they've also open-sourced their code, weights as well as eval pipeline)
Paper:
Code:
@blowdart
"I am glad we agree. What convinced you otherwise in these past two weeks?"
It allows them to present how their mental model works which you could use to have an easy time persuading or reasoning with them next time or counterargue with urself when u want a second opinion.
Finally at the point where I like my website/blog. New release there today 🫣
As July rolls in, will be taking a small break from blog-posts to focus on developing some open-source projects for people interested in learning how to build LLM projects.
Specific requests anyone?
What evaluation frameworks for LLMs are you using or have seen being used aka link me to the blogpost etc?
Asking for our research paper 📜 We are doing a survey paper.
RT for reach plz.
Will post about embeddings - 0 bookmarks, pity likes from friends only
Posts a casual photo of me wearing a sketch girl drinking man tears - 10 bookmarks 😱
Did I pick the wrong field? Maybe should've gone for fashion influencer instead of machine learning engineer 😂
My friend
@DataScienceHarp
's company Deci has launched a new Foundation Model for Object Detection
And the results are 🔥🎇
Here's why you should care about it 🧻 (1/n)
#yoloNAS
#computervision
You have only 14 hours a day to dedicate to your work, one of the most imp skills is to tame your ego and say "no" to a ton of great opportunities every single day.
Focus.
Don't do everything. Do a few things incredibly incredibly well and let compounding do the rest.
While this one by Cote et al is the description of researchers' approach to designing a study that would hopefully guide how to build quality assurance tools in ML Software Systems (study is yet to be out) but it does bring attention to an open challenge.
At 16, had this opportunity to train as a cadet. Dad told battalion head I was too fragile for sun and guns 🤣 Went to another batallion, got selected. Trained hard for 2 years. Came back after a camp with 8 medals and rifle injuries that took 6 months to heal 🤭 Slayed it🤩
Fully agreed with this concern. We are likely heading into a deep AI winter the moment we run into some colossal failures at scale, the probability of which seems incredibly high in the hands of self-interested folks who have likely never deployed ML systems at scale before LLMs.
4. Dalton: Learned Partitioning for Distributed Data Streams addresses skewed data yield imbalanced load assignments as well as scalability issues for high-throughput streams modern stream processing systems.
Right now, it's a little too early to define what will the stack for LLM Engineers or AI Engineers (whatever the name comes out to be) will look like. This is probably amongst one of the variations I have seen out there.
@jerryjliu0
@swyx
Agreed.
I think a lot of conventional MLOps is created around discriminative models whereas LLMs being generative poses a whole new set of challenges that requires a new stack and a different framework for both the kinds of deployments.
God knows how people manage to smile for photos when no one is fooling around 🤣
I can barely manage just a little teeth and my usual puppy dog eyes 🤣😂
Anyway, got a new haircut - not like anyone would notice but short hair are back in business 🤣💃
@yoheinakajima
@geri_kirilova
is one of the rare ones that is actually working on the ground to fix it and doing such a fabulous job with
@ventureco_op
I won't be surprised if Laconia is in 10-15 years where a16z/sequoia is today as best of the best thought-leader in the VC industry.
As we are close to wrapping up our research paper on LLMs, we are next taking a stab at LLM Agents (eg. AutoGPT, BabyAGI etc) - open problems, limitations, framework, use cases in production.
Our kickoff meetings begin on June 18.
Tell me what you're working on I'll DM the link
Had a really good time at the AI Engineer Summit. Time to hit 💤
The day after tomorrow i.e. oct 13, I will be presenting my interpretation of LLMOps for the first time and how to build robust AI/ML systems in production at the virtual conf Put Gen AI to work by Packt.
MLops companies in the future will be more like digital marketing companies. Some will build succesful automation products as did Hootsuite and many more shall serve as the equivalent of digital marketing agencies.
And it just went live. Join me on Wednesday, March 20 with AI Tinkerers - Ottawa
where we talk:
- why aren't we seeing more AI gents in production?
- Is LLM the path to AGI?
- LLMOps vs MLOps
- LLMs on resource-contrained devices
& more
3. Streaming vs. Functions: A Cost Perspective on Cloud Event Processing by Tobias, Soren et al, implement stateless and stateful workflows to benchmark cloud cost for both Function-as-a-Service (FaaS) as well as distributed stream processing (DSP).
Less than 24 hours to go!
RSVP if you haven't 𝗛𝗼𝘄 𝗱𝗼 𝘆𝗼𝘂 𝘁𝗮𝗶𝗹𝗼𝗿 𝗴𝗲𝗻𝗲𝗿𝗶𝗰 𝗟𝗮𝗿𝗴𝗲 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗠𝗼𝗱𝗲𝗹𝘀 𝗳𝗼𝗿 𝘆𝗼𝘂𝗿 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀? with
@AndreasMWelsch
on LinkedIn. Don't forget to add to GCal - see u there!
RSVP 👉👉
We had such a good turnout.
3333 people (yes, really) signed up for the 3 hour online event.
About 650 showed up. More than 85% stayed till the very end.
O'Reilly team are absolute pros when it comes to online events. Such seamless delivery 🫡
AI Live Online Event: Building AI Agents with Large Language Models -- Join us with hosts Nicole Butterfield &
@GoAbiAryan
on Aug 29 12PM EDT. Learn about the latest tools to help you build your own AI agent. Sign in, or sign up for your free trial:
Evaluation is always the hard part with large language models.
The focus with PaLM 2 was multi-linguality and sentiment analysis (remember the chatbots that got hateful and abusive, they tried to mitigate that)
Google Photos sent a notification that they have a cinematic edit on one of my photos from the other day🙈
Wow, wtf, this stuff is trippy 🤣 🤷♀️
No blinking required whatsoever 😉
📜Improving Language Model Negotiation with Self-Play and In-Context Learning from AI Feedback
⁉️if LLMs were able to improve each other, it would imply the possibility of creating strong AI agents with minimal human intervention.
💡Result: they can't