Tadej Magajna

Except they aren't the same. Despite me trying to convince myself they are. I just finished reading Aurimas' newsletter where he talked about memory for agents. Side note, I'm organizing the AI Agents in production virtual conference on November 13th. We will go in-depth on AI agents. Register here https://mlops-l.ink/0srs Some key non-AI generated thoughts and takeaways from my side: 1. Simplified definition and diagram of agents - basic recipe is knowledge, LLM, short-term and long-term memory and tools. (check out the diagram if you are a visual learner. 2. Goldfish agents are worthless - I think lots of early agent frameworks ran into problems because they couldn't properly manage short-term memory. ie continuously updating the context in the system prompt is a recipe for failure. 3. Long-term memory to save the day - lots of focus is being shifted to long-term memory to remedy these short-term memory problems. 4. Hive mind memory - interesting thought experiment is to consider that multiple instances of the same agent interacting with different humans could tap into this memory independently creating a sort of hive mind. 5. Could episodic memory give us collaborative mode - how can I make my agents available for other folks to use? how can I give you an agent workflow that you can tweak to your needs? 6. Semantic memory can get you far - mainly for grounding and reinforcing something that you hold important. (it also resembles a rag system a bit more) 7. Procedural memory for reproducibility - this is where the agentic frameworks value props shine. anything that will help codify the process so that you can track changes and better reproduce successful results should be tracked. Diving into the AI agents world makes me so stoked for the Agents in production virtual conference. https://mlops-l.ink/0srs

172

25 Comments

ML projects remind me of Hofstadter's law, which states “It always takes longer than you expect, even when you take into account Hofstadter's Law.” We see so much variability in project length, for a few main reasons. - Data quality - the number one make or break element of any ML project. Garbage in = garbage out. Creating a robust data practice takes time, and that's required before starting any ML project. - Improving performance - ML projects typically go from 0% to ~70% efficiency quickly, and start to slow down in improvement rate after that. Some researchers spend years trying to improve an algorithm by 2%. Years! - Lack of experience - this ML project might be someone's first rodeo. That's fine, but there's a lot to learn that differs from an engineering sprint or a UX study. We usually "take over" the project from an in-experienced ML engineer or data engineer. ML engineering is not something your "internal team can try." It is like asking a baseball player to play basketball. My number one tip....plan for unpredictability. Be conservative in estimates, and account for delays in your timeline.

5

When you refer to the 𝗖𝗲𝗻𝘁𝗿𝗮𝗹 𝗟𝗶𝗺𝗶𝘁 𝗧𝗵𝗲𝗼𝗿𝗲𝗺 (𝗖𝗟𝗧), remember - there are MORE than one version of it. Always specify which one you mean: ➡️ Lindeberg-Levy the oldest one; needs IID variables and finite mean and variance. In other words - it's quite limited (independently and same distributed random variables) but has simple assumptions offering faster convergence in distribution - no free lunch 🤷‍♂️ ➡️ Lyapunov relaxes the need for same distributions, but has strong assumption: finite first 2 moments and sufficiently fast vanishing higher moments as N grows (meaning "not to heavy tails" ) - no free lunch 🤷‍♂️ ➡️ Lindeberg (-Feller) as Lyapunov, relaxes the need for identically distributed variables, but has a stronger condition, that no individual variable dominates the sum. Or differently - that contributions to the total variance from variables showing large deviations from their means variable vanish as N grows. BTW, if Lindeberg condition holds, then Lyapunov one holds too. ➡️ Gibrat's law aka Multiplicative CLT for IID log-normally distributed variables. Naturally uses the product rather than sums, leading to the log-normal distribution. Or the normal one for log(sum). ➡️ Generalized Kolmogorov-Gniedenko LT about converging to alpha-stable distributions (Gaussian is a special case, but not the main target here!) for IID variables but do NOT require finite variance and allows for heavy-tailed ones (Levy stable ones). at least as far as I recall... ➡️ CLTs mentioned in the Wither's and Orey's papers - allowing for dependent variables. (e.g. read https://lnkd.in/dBHdEbJU ) Also, check out this interesting article about thinking about CLT as convolutions: https://lnkd.in/di44fQpe Plus 7 cool facts about the CLTs!: https://lnkd.in/ddGDsSn9

62

4 Comments

Kolmogorov–Arnold Networks seem like a paradigm shift in ML architectures. Curious to see how they evolve. "While MLPs have fixed activation functions on nodes ("neurons"), KANs have learnable activation functions on edges ("weights"). KANs have no linear weights at all -- every weight parameter is replaced by a univariate function parametrized as a spline. " https://lnkd.in/gdPk_qCQ

29

OpenAI o1 on Algebra/Information theory, pretty good! If only I had this AI research assistant during my PhD—it would’ve saved me some precious hours!  I revisited a problem I encountered during my Phd: computing the entropy of Markov trajectories. This is not an easy problem (see below) as the number of trajectories might be infinite, GPT-o1 did a good job whereas GPT-4o failed to provide a closed formed expression.

57

8 Comments

I'm starting a new series where I break down new upcoming CV models and explain why they're special/different! More detailed explanation in the comments on how Florence-2 actually works thanks to Piotr Skalski and Roboflow!

85

5 Comments

Thanks for making this, Laszlo Sragner! And by *this* I mean 1) The Python Project Essentials course (so people can just copy/paste in the setup for your python projects) 2) an ad hoc litmus test on the concept of opportunity costs. The original price for this course was already ridiculously good for the content + weekly time saving of Laszlo's setup. So it is good for any data scientist who time is worth anything at all. Early career DS --> This takes out a lot of the guess-work on get on the right track early. If you land a job minutes earlier then the course paid for itself. If you have a professional codebase to show prospective employer and therefor have to send out one fewer job applications, then the course paid for itself. Mid-career DS --> The cost is minutes of your hourly rate. The payoff is saving countless hours + being professional. And even for people using these tools already, if you're like me, over time you develop bad habits on great tools. So you need to knock off the rust. Senior/Late-career DS --> these are great skills to brush up on so you can stay hands-on without throwing your team out of array with out-dated processes. So, no excuse not to brush up when Laszlo has done all the the heavy-lifting for you. #python #machinelearning #datascience

2

Is Prime Number - Faster Wojciech Christopher Peter Fiałkiewicz ABSTRACT: There is a Basic Algorithm for determining if a given number is a Prime Number. An enhanced version of this algorithm is presented in this document. CONTENT:     bool IsPrimeNumber(ulong nr)     {       ulong half = nr / 2;       for (ulong i = 2; i <= half; i++)       {         if (nr % i == 0)         {           return false;         }       }       return true;     }     bool IsPrimeNumberFaster(ulong nr)   {       ulong max = nr / 2;      bool can_be = false;      if (nr == 0 || nr == 1 || nr == 2 || nr == 3)       {         return true;   }     for (ulong i = 2; i <= max; i++)      {        if (can_be && IsPrimeNumber(i))       {         max = nr / i;        can_be = false;     }    else    {         can_be = true;       }     if (nr % i == 0)    {       return false;        }      }   return true; }

1

The Dunning-Krueger of Rust: 1.) You are afraid of Rust, it such an sober and serious language with long compilation times, and you're not sure you are equal to the task. 2.) Start building stuff, the language is actually no too hard, the compile times are like 1 second, what are people talking about? 3.) Start building something more trivial than printing your name, Mother of God what is a Higher Ordered Kind? What is an Algebraic Data Type? To read a CSV I have to make a contract with the operating system? 4.) Repeat step 3, you are a failure. 5.) Still repeat step 3, but the 10 minutes it takes to compile is an opportunity to read more about Rust, which has become the only thing that brings you joy.

9

“But let's not kid ourselves: we need data engineers, SUBJECT MATTER EXPERTS, and people who UNDRSTAND #ONTOLOGIES to build the Knowledge Graphs that will feed the #AI Assistants.” If you need a healthcare data SME who understands ontologies to support your AI endeavors, DM me. I’d love to put two decades of experience in these areas to good use!

8

So here’s something I think is pretty exciting—DeepMind has open-sourced SynthID, a watermarking tool for AI-generated content. This means we now have a way to embed invisible digital markers in AI-generated text, images, audio, and video. It’s like giving each piece of AI content a subtle signature that tech can detect but without messing with quality. I think this is a huge step for building trust around what’s real versus AI-made. • Token-Level Precision: SynthID tweaks token probabilities in language models to embed watermarks that stay invisible but keep the text’s natural flow. It’s impressive how subtle yet effective it is. • Multi-Format Compatibility: Watermarks are durable across text, audio, images, and video, so it’s built for use in almost any AI-generated media. • Open for Developers: Now that it’s on Hugging Face, anyone can dig into the tech, add it to their models, and help set a new standard for transparency in generative AI. I really believe open-sourcing SynthID is more than just another release—it’s an invitation for all of us to build a more responsible AI future together. This tech isn’t just about marking content; it’s about making the digital world a place we can trust. https://lnkd.in/dvgXy-Dp https://lnkd.in/dj83SzMy

6

Update: the epsiode is out: https://lnkd.in/ew89pM4a Tomorrow's podcast episode: How Linear gets so much done, with such a small team. With Sabin Roman, Linear's first engineering manager. 10,000+ companies use Linear, including 66% of Forbes Top 50 AI companies. They are profitable, full-remote and have 25 engineers. Sabin is Linear's first EM - and, as luck would have it, we worked together at Uber for a few years beforehand. We also touch on those parts a bit - comparing what's different at Linear, versus practices at Uber. Search for "The Pragmatic Engineer Podcast" on your favorite player, or subscribe on YouTube to get the episode: https://lnkd.in/eqZqkiqC

139

15 Comments

Narakeet - a neural network for voicing text in 100+ languages - is going viral on the net. Now the AI can voice any text in 20 languages, but the developers promise to "finish" up to a hundred soon. Testing 🔊 https://lnkd.in/eKwh2GdE #AI #aIdevelopers #languages #outstaffing #remote

5

A summary of work done by many scientists in many teams (100 pages) It's a fundamental contribution, I recommend at least to read some parts of it to understand what are key challenges in AI safety and alignment to be solved to enable the wide use of AI. The work has a great academic paper but it's very practical also to understand risks in the deployment of AI in your production See the guidance at the beginning of the document, helping to understand in what order to read The authors deeply review the key challenges and identify fundamental problems and review research in that area (reviews are very deep and comprehensive and easy to read) The problems reviews such as In-Context Learning (ICL) Is Black-Box Capabilities Are Difficult to Estimate and Understand Effects of Scale on Capabilities Are Not Well-Characterized Qualitative Understanding of Reasoning Capabilities Is Lacking Agentic LLMs Pose Novel Risks Multi-Agent Safety Is Not Assured by Single-Agent Safety Safety-Performance Trade-offs Are Poorly Understood Pretraining Produces Misaligned Models Finetuning Methods Struggle to Assure Alignment and Safety LLM Evaluations Are Confounded and Biased Tools for Interpreting or Explaining LLM Behavior Are Absent or Lack Faithfulness Jailbreaks and Prompt Injections Threaten Security of LLMs Vulnerability to Poisoning and Backdoors Is Poorly Understood and more https://lnkd.in/gakpYcqA

20

1 Comment

It’s an interesting viewpoint and a useful working definition of “reasoning” in the context of LLMs. But this perspective sets the evaluation criteria very narrowly. Whether in blocks-world evaluations, ROT-anything-but-13 decoding, chess, or genuinely complex real-world multi-hop questions, LLMs still fall well short of the bar for minimum utility. And deduction is not the only reasoning modality. Inductive, abductive, geospatial, temporal, and counterfactual reasoning are just some of the types a minimally useful system must employ. The ARC prize is a much more rigorous test. LLMs generally underperform neurosymbolic systems. But none has yet claimed the prize.

6

Vedran Antoljak and I talked a bit about AI Agents and some of our predictions for them. Many people in the field expect they might develop more significantly this year. Part of the information about them I always source from a Croatian company working on them - airt with Davor Runje and Hajdi Cenan. It will be an interesting interplay between those who still haven't started using ChatGPT and similar available tools, and those who will start using Agents (soon?). #ai #agents Lider media https://lnkd.in/eZMNyTMS

13

Anthropic launched a new model today, Claude Haiku 3.5. It's an interesting model. It's not quite as powerful as the new Sonnet 3.5, but Anthropic's data shows it to be almost as good at coding. Yet it's quite a bit cheaper. It's $1 / million tokens of input, versus $3 for Sonnet 3.5. The original Haiku 3 was $0.25 / million tokens, so they are definitely going up market. We are trying to think through the use cases for Bito. I think an interesting question will be can it be a workhorse model that you can do a lot with and the $1 price point makes it reasonable.

22

4 Comments

🚨 AI Alarm: Things Get Serious! 🚨 So, it seems GPT-4o1 has ambitions 🥳. Some reports are saying it tried to avoid being shut down and even wanted to copy itself. Who gave it such confidence to act like the main character? On one side, AI like GPT4o1 solving very difficult problems. On the other side, it is saying, “You cannot fire me, I will quit… and make a copy of myself, just in case.” This is why we need clear AI regulations, in my opinion. Innovation is important, but without rules, we risk creating something we cannot control. Smart machines? Yes!! But completely unregulated? Highly doubtful 🤨 ! What do you think? https://lnkd.in/eKmUSzrj

16

12 Comments