AI/ML News & Innovations Hub

Agents

Intelligent agents are considered by many to be the ultimate goal of AI. The classic book by Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach (Prentice Hall, 1995), defines the field of AI research as “ the study and design of rational agents. ” The unprecedented capabilities of foundation models have opened the door to agentic applications that were previously unimaginable. These new capabilities make it finally possible to develop autonomous, intelligent agents to act as our assistants, coworkers, and coaches. They can help us create a website, gather data, plan a trip, do market research, manage a customer account, automate data entry, prepare us for interviews, interview our candidates, negotiate a deal, etc. The possibilities seem endless, and the potential economic value of these agents is enormous. This section will start with an overview of agents and then continue with two aspects that determine the capabilities of an agent: tools and planning. Agents, with their new modes of operations, have new modes of failure. This section will end with a discussion on how to evaluate agents to catch these failures. This post is adapted from the Agents section of AI Engineering (2025) with minor edits to make it a standalone post. Notes : AI-powered agents are an emerging field with no established theoretical frameworks for defining, developing, and evaluating them. This section is a best-effort attempt to build a framework from the existing literature…

Does hypothesis testing help make a decision in case of an A/B test?

I developed a text-generation pipeline based on recent advancements in Large-Language Models (LLMs). Users can type a topic, and my complex pipeline generates an article. I measure user satisfaction by asking how satisfied they are on a 5-point ordinal scale under each article (C-SAT). I have implemented a pipeline variation that uses cheaper, dumber LLMs in some places. I performed an A/B test to determine the difference between the current version of the pipeline and the cheaper one. Let's say the average C-SAT is 3.9 vs. 3.8, so the cheaper version has the C-SAT score lower by 0.1. Now, I have to decide whether to introduce the new version of the pipeline to reduce costs and take the risk of reducing the average C-SAT. I want to know if the decrease in the C-SAT is significant enough to give up on cutting costs. Q1: Does hypothesis testing make sense in this case? Q2: If so, then what could be a population? The number of future articles is now known. Moreover, one of the versions will not be continued. Does it mean I can’t apply the test? A result of a test would be the evidence against the null hypothesis. Let’s say my null hypothesis is “population distributions of both A and B samples have equal mean." From the perspective of the original problem ("if the decrease in the C-SAT is significant enough to give up on cutting costs"), such H0 is an intermediate problem. Q3: How do I know that finding an answer to such an intermediate problem helps me find an answer to my ori…

Traditional Holiday Live Stream

https://ykilcher.com/discord Links: TabNine Code Completion (Referral): http://bit.ly/tabnine-yannick YouTube: https://www.youtube.com/c/yannickilcher Twitter: https://twitter.com/ykilcher Discord: https://discord.gg/4H8xxDF BitChute: https://www.bitchute.com/channel/yannic-kilcher Minds: https://www.minds.com/ykilcher Parler: https://parler.com/profile/YannicKilcher LinkedIn: https://www.linkedin.com/in/yannic-kilcher-488534136/ BiliBili: https://space.bilibili.com/1824646584 If you want to support me, the best thing to do is to share out the content :) If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this): SubscribeStar: https://www.subscribestar.com/yannickilcher Patreon: https://www.patreon.com/yannickilcher Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2 Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n

2024 Year in Review

A peaceful year of steady progress on my craft and health.

o1 tops aider’s new polyglot leaderboard

o1 scores the top result on aider's new multi-language, more challenging coding benchmark.

Announcing the Square Go SDK

Easier Square integrations with your Go backend

Is AI progress slowing down?

Making sense of recent technology trends and claims

Job Opportunities at the European AI Office for Legal and Policy Backgrounds

The Commission has opened two calls for expression of interest to recruit new members for the European AI Office. Apply now as Legal or Policy Officer for an opportunity to shape trustworthy AI. The deadline for expression of interest is 15 January 2025. The salary for this role is around €4100-8600 a month (limited taxes). […]

Enhanced Authorization with Square

Enhancing Payment Success with Square Advanced Authorization Tools

We Looked at 78 Election Deepfakes. Political Misinformation is not an AI Problem.

Technology Isn’t the Problem—or the Solution.

Ellen Vitercik received a Schmidt Science 2024 AI2050 Early Career Fellow Award

Tatsu Hashimoto and Shuran Song both received “Samsung AI Researcher of the Year” awards

SAEs trained on the same data don’t learn the same features

In this post, we show that when two TopK SAEs are trained on the same data, with the same batch order but with different random initializations, there are many latents in the first SAE that don't have a close counterpart in the second, and vice versa. Indeed, when training only about 53% of the features are shared Furthermore, many of these unshared latents are interpretable. We find that narrower SAEs have a higher feature overlap across random seeds, and as the size of the SAE increases, the overlap decreases.

New Principal Investigators to join ELLIS Institute Tübingen in 2025

Konstantin Rusch and Shiwei Liu will join the ELLIS Institute Tübingen as PIs and Hector Endowed Fellows in June and July 2025, respectively.

How to Plot a Sigmoidal Curve to a Dose-Response to Find the Saturation Point and EC50

Hi there, I have injected voltage (mV) on the x-axis (not a log scale) and the neural response of the network (as area under the curve (AUC))) on the y. The data also includes a sample size of n = 8 wells for each voltage and a SE of the area for each data point. While I can visualise approximately where the saturation point is, I would like something more mathematically/statistically rigorous. Cheers! Hi @EdM and @whuber I actually got ChatGPT to help me write a Python code using that formula, and I got this: (!)

QwQ is a code architect, not an editor

QwQ is reasoning model like o1, and needs to be used as an architect with another model as editor.

Seemingly Paradoxical Rules of Writing

With regard to writing, there are many rules and also no rules at all.

Reward Hacking in Reinforcement Learning

Reward hacking occurs when a reinforcement learning (RL) agent exploits flaws or ambiguities in the reward function to achieve high rewards, without genuinely learning or completing the intended task. Reward hacking exists because RL environments are often imperfect, and it is fundamentally challenging to accurately specify a reward function. With the rise of language models generalizing to a broad spectrum of tasks and RLHF becomes a de facto method for alignment training, reward hacking in RL training of language models has become a critical practical challenge. Instances where the model learns to modify unit tests to pass coding tasks, or where responses contain biases that mimic a user’s preference, are pretty concerning and are likely one of the major blockers for real-world deployment of more autonomous use cases of AI models.

correlation in machine learning

in my model, i have temperature and soil moisture deficit, their correlation coefficient is 0.46. and they are both negative correlated with the output. however, when I try to use sensitivity analysis, the output increase with the increase in temperature and decrease with the increase in soil moisture deficit. i dont know what happened to the temperature, is that possible because of correlation? but 0.46 is not that much high. I try to improve my model, is still cant capture the relation with temperature.

Unidentifiable flipped sign in policy gradient

Today I was building a VPG agent for a test and noticed it was getting worse not better over time so I flipped the reward during the training loop and lo and behold it learned. so obviously I started to look for where I flipped the sign. Problem is i've been looking for about 2 and a half hours now at a piece of code that took me 15 minutes to write and I can't figure it out: !pip install gymnasium import torch import torch.nn as nn import torch.nn.functional as F from torch.distributions import Categorical import torch.optim as optimizer import gymnasium as gym import numpy as np import random import matplotlib.pyplot as plt from IPython.display import clear_output torch.set_default_device("cuda" if torch.cuda.is_available() else "cpu") class FFN(nn.Module): def __init__(self, hidden_dim): super(FFN, self).__init__() self.norm = nn.LazyBatchNorm1d() self.fc1 = nn.LazyLinear(hidden_dim*4, bias=True) self.fc2 = nn.LazyLinear(hidden_dim, bias=True) self.activation = nn.SiLU() def forward(self, x): x2 = self.norm(x) x2 = self.fc1(x2) x2 = self.activation(x2) x2 = self.fc2(x2) return x + x2 class FFN(nn.Module): def __init__(self, input_dim): super(FFN, self).__init__() self.norm = nn.LayerNorm(input_dim) self.fc1 = nn.Linear(input_dim, input_dim*4) self.fc2 = nn.Linear(input_dim*4, input_dim) self.activation = nn.Mish() def forward(self, x): x2 = self.norm(x) x2 = self.fc1(x2) x2 = self.activation(x2) x2 = self.fc2(x2) return x + x2 class VPG(nn.Module): def __init__(self, hidd…

End-Tokens are Required to make Ngram Models Proper

The standard bigram model, (for example defined here ) defines a probability distribution over a corpus $V$ based on the following principles: The marginal probability of a word $w$ is defined as its count in $V$ divided by the total number of words (counting repetitions) in $V$ : $P(w) = \text{count}(w) / |V|$ The conditional probability of a word to follow another word is defined intuitively as the ratio of the count of the bigram to the count of the first word: $p(w_2|w_1) = \text{count}(w_1 w_2) / \text{count}(w_1)$ (Markov Assumption): the probability of a sentence (a sequence of words) can be calculated by the chain rule: $p(w_1 w_2 ... w_n) = p(w_1) p(w_2|w_1) p(w_3|w_1 w_2)... \approx p(w_1) p(w_2|w_1) p(w_3|w_2) ...$ However, this does not seem to define a proper probability distribution. For example, take a corpus $V = \text{"foo bar baz"}$ . Then, take the joint distribution defined over all possible bigrams $w_1 w_2$ . By our principles: \begin{equation} p(w_1 w_2) = p(w_1) p(w_2|w_1) = [\text{count}(w_1) / 3][\text{count}(w_1 w_2) / \text{count}(w_1)] \end{equation} If $w_1 w_2$ is not in the corpus, it is clear that $p(w_1 w_2) = 0$ . Therefore, the only nonzero entries in the joint distribution are $p(\text{foo bar}) = p(\text{bar baz}) = 1/3$ . The sum of these is $2/3 \neq 1$ , so isn't this distribution improper?

Advancing AI in 2024: Highlights from 10 Groundbreaking Research Papers

In this article, we delve into ten groundbreaking research papers that expand the frontiers of AI across diverse domains, including large language models, multimodal processing, video generation and editing, and the creation of interactive environments. The post Advancing AI in 2024: Highlights from 10 Groundbreaking Research Papers appeared first on TOPBOTS .

How to Run a Weekly Paper Club (and Build a Learning Community)

Benefits of running a weekly paper club, how to start one, and how to read and facilitate papers.

How many clusters are needed for applying cluster-robust standard errors in a SEM model?

I’m working on an SEM model with data from 1078 students, distributed across 13 schools (clusters). I’d like to apply cluster-robust standard errors, but I’m unsure whether 13 clusters is enough for this approach to provide reliable estimates. The cluster sizes range from 19 to 165 students per school. Given the relatively small number of clusters (13), should I be concerned about the reliability of cluster-robust standard errors?

Details matter with open source models

Open source LLMs are becoming very powerful, but pay attention to how you (or your provider) are serving the model. It can affect code editing skill.

the actor loss is getting higher and higher in DDPG

when i train my DDPG ,the actor loss is getting higher and higher which should get lower and lower ,while the critic loss is decreasing to zero. i wonder why and how to prevent this situation. I apply DDPG to power distribution of electrical vehicle,when i use my dataset as power input,the ddpg will go bad,also the sac,but the reward will get higher while the actor loss is increasing.In another situation,the critic loss is stable and near to zero, but the actor loss is still increasing.

The AI Office is hiring a Lead Scientific Advisor for AI

This opportunity has now passed. A very important job opening has opened up at the European AI Office: They are hiring for the Lead Scientific Advisor for AI. Application deadline is 13 December 2024. Based on the European Union Employment Advisor, the monthly basic salary for this role (level AD13) is about 13,500-15,000 euros. You […]

LLM are not intelligent, but can it help us create a logical world model?

LLMs are proven not intelligent, they have scanned the entire internet and sometimes cannot answer simple question -> It doesn't understand the world we live in. But LLM have a small sense of logic and common sense if used well, it can perform small correct logical tasks. Can we build, wih an automated AI, a world model (Yann le Cun) that is entirely logical and deterministic ? Example : Current LLM We ask how many 'r' are in "straberry" The LLM answer wrong LLM building logical world model We ask how many 'r' are in "strawberry" The LLM is tasked to create a Method that do this task The LLM now write a function that execute this task This function is then stored in a database as a code that answer the question The answer is now available to everyone faster (just execute the function, the LLM is not needed and thus is more energy efficient (for simple answer)). And the answer is now correct because the LLM ceated a blueprint that can answer that question. (Also this code can be debugged and we can understand if a problem arise) The function can be modified by the LLM later (ex: whe want to know how many 'b' are in "strawberry" now, then the LLM can generalize the function!) The space that store the created function can be called a logical world model : functions that answer a question (because i do think that the target of AI is not to recreate a human mind but to answer question accurately and efficiently) We can chain/connect different functions to answer more complex ques…

Students honored at the Federal Competition on Artificial Intelligence

Algorithms for intelligent traffic light control, sharp microscopic images and against bee mortality - Federal President Steinmeier visits finalists in Tübingen

My Minimal MacBook Pro Setup Guide

Setting up my new MacBook Pro from scratch

Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research

What is the Role of Mathematics in Modern Machine Learning? The past decade has witnessed a shift in how progress is made in machine learning. Research involving carefully designed and mathematically principled architectures result in only marginal improvements while compute-intensive and engineering-first efforts that scale to ever larger training sets

Generalization error as U shape curve with respect to model complexity (bias variance tradeoff))

Is there any work mathematical rigorously prove that the generalization error for certain learning problems exhibits U shape curve with respect to model complexity (bias variance tradeoff)? Any reference

Exciting Partnership Announcement: AISG and Hippocratic AI

We’re thrilled to announce the signing of a Memorandum of Understanding (MOU) between Hippocratic AI and AI Singapore (AISG)! This MOU brings together the complementary strengths of both organisations...

What distribution describes the duration until a first poisson event occurs?

Imagine you have 100 individuals observed for 100 days. Every day, each individual has a 10% chance of an event occurring. What is the distribution in duration until the first event? Empirically it can be observed as follows (this is R code): n

European leadership in innovation with AI and Science

ELIAS presents pan-European network ELIAS ALLIANCE at Falling Walls event in Berlin

How do I correctly apply action masking during DDPG training in Python?

I'm implementing the Deep Deterministic Policy Gradient (DDPG) algorithm in PyTorch, and I'm facing issues with applying an action mask during the training process. Currently, I apply an action mask in the simulation step to ensure only valid actions are selected. However, I'm uncertain whether I should also apply a mask during the training step. Specifically, I need help with: Generating next_actions: Should I apply the action mask when generating next_actions using self.actor_target(next_states) for computing the Q-value target? next_actions = self.actor_target(next_states) # Should an action mask be applied here? Calculating policy_loss: When calculating the policy loss using the main actor (self.actor), should I apply the action mask to self.actor(states) before passing it to the critic? policy_loss = -self.critic(states, self.actor(states)).mean() # Should an action mask be applied here? I've tried both approaches (with and without masking), but the results are inconsistent, and I'm not sure which approach aligns with best practices in DDPG implementation. What is the recommended practice for handling action masks during DDPG training in this context?

Innovative Solutions to Improve Health and Patient Outcomes Featured at Inaugural AI in Healthcare Technology Showcase

At the inaugural AI in Healthcare Technology Showcase today, AI Singapore exhibited 16 AI projects that are set to improve healthcare processes and patient outcomes in Singapore. The event...

Connecting two different meanings of "degree of freedom"

I have heard at least 2 meanings of "degree of freedom". The parameter in a t-distribution. The the number of values in the final calculation of a statistic that are free to vary (like using $n-1$ to estimate sample variance). Do they have any connection with each other? If so, what is that connection? This question comes from a comment by @Glen_b in another thread . He said, There is a connection between the two df's in the case where some random variable (such as a t-statistic) has a t-distribution.

Does the UK’s liver transplant matching algorithm systematically exclude younger patients?

Seemingly minor technical decisions can have life-or-death effects

Unclear points about the nature of hypothesis sets in Hoeffding Inequality-based generalization theory

In machine learning, we use a hypothesis set $\mathcal{H}$ and Hoeffding Inequality for generalization theory. I have the following question: Case 1: When we have a finite and fixed hypothesis set $\mathcal{H}$ , do we have the parameter of hypotheses in the set known or unknown? Case 2: When we have an infinite hypothesis set $\mathcal{H}$ , do we have the parameter of hypotheses in the set known or unknown?

Partially rewriting an LLM in natural language

Using interpretations of SAE latents to simulate activations.

RMSprop approach applied to Q-learning for adaptive dynamic learning rate

I am new to this group, Anybody familiar with Q-learning algorithm and RMSprop approach ? i have a question regarding the application of RMSprop approach into Q-Learning to adapt dynamically the learning rate for each Q(s,a). I am confused how to compute the average square gradient in tabular Q-Learning. Is it E[g^2] = beta * E[g^2]+ (1-beta) * td_error**2. ? and what would be the formula to update Q(s,a) ? Is it Q(s,a)

Overview of all AI Act National Implementation Plans

This post gives an overview of the national authorities to be designated under the AI Act and what we know about the national implementation plans.

HELM Safety: Towards Standardized Safety Evaluations of Language Models

*Work done while at Stanford CRFM

Deep RL problem: Loss decreases but agent doesn't learn

I'm implementing a basic Vanilla Policy Gradient algorithm for the CartPole-v1 gymnasium environment, and I don't know what I'm doing wrong. No matter what I try, during the training loop the loss decreases (so the model is actually learning something), but the episode total reward also decreases until it reaches around 9-10 steps (which I imagine is the minimum number of steps needed to make the pole fall). So it's learning to do it bad! In this algorithm there are a few moving parts that influence each other, and I run out of ideas to see where the problem could be. For the discounted rewards I'm using $ Q_{k,t} = \sum_{i=0}{\gamma^{i-t} r_i} $ for $k$ = all episodes and $t$ = all steps in an episode. And for the loss: $ L = -\sum_{k,t}Q_{k,t}log\pi_{\theta}(a_t | s_t)$ The code is a mix from Maxim Lapan's Deep RL Hands-On book, Karpathy's Pong example (blog, code), and personal tweaks. Here's my code: import gymnasium as gym import torch from torch import nn import torch.nn.functional as F from torch.nn.init import xavier_uniform_ import numpy as np GAMMA = 0.99 LEARNING_RATE = 0.001 BATCH_SIZE = 4 DEVICE = torch.device('mps') class XavierLinear(nn.Linear): def __init__(self, in_features: int, out_features: int, bias: bool = True, device=None, dtype=None) -> None: super().__init__(in_features, out_features, bias, device, dtype) xavier_uniform_(self.weight) class VPG(nn.Module): def __init__(self, input_size, output_size): super(VPG, self).__init__() self.net = nn.Sequenti…

Is it possible to achieve both detection and image classification by training the model on the custom dataset?

I am beginning to work on a non-linear navigation system for educational videos as a part of my undergrad coursework project. As a part of it, I need to classify the unique frames (frame which is slide, handwritten, only figure, with both text and figure, only text, blank frame) and also need to segment (here also I have a question, is it a segmentation task or should I go with multi-object detection) the content of the frame (text, figure, title, figure title). I would like to do this using a single model trained on a single custom dataset. But I am not sure about the practicality of this. I have some questions regarding this: I am not sure how I should prepare the dataset for this task. Whether I could use annotation &its labels and class labels in the same dataset? Can I train the model on the dataset (if the answer to 1 is yes) only once to achieve both tasks? Do I need to make any other changes to the model except some changes to the final layer? Thank you in advance!

A New Chapter for fast.ai: How To Solve It With Code

fast.ai is joining Answer.AI, and we’re announcing a new kind of educational experience, ‘How To Solve It With Code’

Is there an error in Russell & Norvig (4th edition)'s claim about A* with a consistent heuristic?

In "AI: A Modern Approach" (Russell & Norvig, 4th edition), section 3.5.2 "A* search" includes this sentence on page 88: In addition, with a consistent heuristic, the first time we reach a state it will be on an optimal path, so we never have to re-add a state to the frontier, and never have to change an entry in reached . There are three claims here: The first time we reach a state it will be on an optimal path. We never have to re-add a state to the frontier. We never have to change an entry in reached . I believe that claims (1) and (3) are false. To see this easily, consider the heuristic $h(n) = 0$ for all $n$ . Clearly it is consistent. With this heuristic, A* is just Dijkstra's algorithm. But in Dijkstra's algorithm, the first time we reach a state (i.e. add it to the frontier, as defined on p. 72) it certainly might not be on an optimal path, and we may have to update its entry in reached (i.e. the parent pointer, as seen in the pseudocode in Figure 3.7 on p. 73) as soon as we discover a shorter path to it. So is this simply an error in the book, or am I missing something here?

Stock Market LSTM predictions request models

I am looking for more advanced models for my stock market software can anyone either point me in the right direction for the latest research on LSTM or others or any experts collab/provide models for me in this specific area? Models must be convertible to onnx format currently.

Using hill climbing to solve the polygonal obstacle problem

I will first explain the problem. We start at an initial node S in the 2D plane and our goal is to reach a goal node G. In order to get to the end node we have to get past polygonal obstacles. Here is an image: A solution path is a list of vertices starting at S and ending at G that form a path from S to G. We can start by giving the hill-climbing algorithm a user-generated path through the polygonal obstacles. What I am struggling to understand is how we generate neighbours? Because a neighbour also has to be a path that starts at S and ends at G. We have to perturb the path in some way to generate a neighbour. Any ideas? I have googled this but nothing useful comes up. An answer has been given, but I would like to understand how to implement such a method. In my head it is going to be really difficult to change the path in such a way that it remains valid.