deep learning - TechTalks

Are we at the cusp of a new era for artificial intelligence?

Ben Dickson — Mon, 21 Apr 2025 12:50:14 +0000

The "Era of Experience" envisions AI's evolution beyond human data, emphasizing self-learning from real-world interactions. But challenges loom for this vision.

The post Are we at the cusp of a new era for artificial intelligence? first appeared on TechTalks.

Everything you need to know about Grok-3

Ben Dickson — Thu, 20 Feb 2025 14:54:29 +0000

Grok-3 storms the AI scene, boasting superior capabilities and competitive benchmarks. Here's everything to know about this new LLM and LRM from xAI.

The post Everything you need to know about Grok-3 first appeared on TechTalks.

New training paradigm prevents machine learning models from learning spurious correlations

Ben Dickson — Mon, 20 Jan 2025 14:26:35 +0000

Meta researchers show how memorization-aware training can help machine learning models avoid developing dangerous biases.

The post New training paradigm prevents machine learning models from learning spurious correlations first appeared on TechTalks.

GEAR turbo-charges LLMs with advanced graph-based RAG capabilities

Ben Dickson — Mon, 13 Jan 2025 20:44:56 +0000

GEAR enhances RAG by automatically extracting triples and using beam search to create and iterate over graph representations from retrieved documents.

The post GEAR turbo-charges LLMs with advanced graph-based RAG capabilities first appeared on TechTalks.

How treating LLMs as “actors” can produce better results

Ben Dickson — Mon, 25 Nov 2024 13:56:08 +0000

Think of LLMs as actors, prompts as scripts, and LLM outputs as performances.

The post How treating LLMs as “actors” can produce better results first appeared on TechTalks.

Mistral expands its reach in the SLM space with Ministral models

Ben Dickson — Wed, 16 Oct 2024 20:10:07 +0000

The new Ministral models outperforms other small language models, including Gemma 2, Phi 3.5, and Llama 3.2.

The post Mistral expands its reach in the SLM space with Ministral models first appeared on TechTalks.

Meta SAM 2 is the most impressive object segmentation model

Ben Dickson — Mon, 05 Aug 2024 07:05:33 +0000

Meta's new object segmentation model, SAM 2, provides near-real-time inference on a wide variety of objects and environments.

The post Meta SAM 2 is the most impressive object segmentation model first appeared on TechTalks.

Energy-Based World Models bring human-like cognition to AI

Ben Dickson — Mon, 24 Jun 2024 12:12:15 +0000

Energy-based world models (EBWM) enable AI systems to reflect on their predictions and achieve human-like cognitive abilities missing in autoregressive models.

The post Energy-Based World Models bring human-like cognition to AI first appeared on TechTalks.

Self-assembling neural networks can open new directions for AI research

Ben Dickson — Mon, 13 Nov 2023 14:00:00 +0000

A new software architecture uses neural development programs (NDP) to self-assemble deep learning models from basic units, like their biological counterparts.

The post Self-assembling neural networks can open new directions for AI research first appeared on TechTalks.

A simple guide to gradient descent in machine learning

Ben Dickson — Mon, 31 Jul 2023 13:00:00 +0000

Gradient descent is the main technique for training machine learning and deep learning models. Read all about it.

The post A simple guide to gradient descent in machine learning first appeared on TechTalks.

The complete guide to LLM fine-tuning

Ben Dickson — Mon, 10 Jul 2023 13:00:00 +0000

Everything to know about LLM fine-tuning, supervised fine-tuning, reinforcement learning from human feedback (RLHF), and parameter-efficient fine-tuning (PEFT)

The post The complete guide to LLM fine-tuning first appeared on TechTalks.

How to teach AI to imitate human thought and action

Ben Dickson — Mon, 03 Jul 2023 13:00:00 +0000

A new technique called "Thought Cloning" trains AI systems on both behavior and reasoning data, making them robust and interpretable.

The post How to teach AI to imitate human thought and action first appeared on TechTalks.

What is low-rank adaptation (LoRA)?

Ben Dickson — Mon, 22 May 2023 13:00:00 +0000

Low-rank adaptation (LoRA) is a technique that cuts the costs of fine-tuning large language models (LLM) to a fraction of its actual figure.

The post What is low-rank adaptation (LoRA)? first appeared on TechTalks.

Are the emergent abilities of LLMs like GPT-4 a mirage?

Ben Dickson — Wed, 17 May 2023 13:00:00 +0000

A new study byStanford University suggests that the emergent abilities of large language models (LLM) are caused by a poor choice of evaluation metrics.

The post Are the emergent abilities of LLMs like GPT-4 a mirage? first appeared on TechTalks.

How to customize LLMs like ChatGPT with your own data and documents

Ben Dickson — Mon, 01 May 2023 13:00:00 +0000

ChatGPT and other LLMs are limited to their training data. Here's how you can customize them with embeddings and your own documents.

The post How to customize LLMs like ChatGPT with your own data and documents first appeared on TechTalks.

What we learned from the deep learning revolution

Ben Dickson — Mon, 10 Apr 2023 13:00:00 +0000

Neuroscientist Terry Sejnowski discusses the early struggles of deep learning, its explosion into the mainstream, and the lessons learned from decades of research and development.

The post What we learned from the deep learning revolution first appeared on TechTalks.

What to know about augmented language models

Ben Dickson — Mon, 03 Apr 2023 13:00:00 +0000

Large language models suffer from fundamental problems, such as failing at math and reasoning. Augmented language models address some of these problems.

The post What to know about augmented language models first appeared on TechTalks.

Microsoft and OpenAI get ahead in the LLM competition

Ben Dickson — Tue, 28 Mar 2023 13:00:00 +0000

Microsoft and OpenAI released a bunch of good new LLM products. Google has released Bard, but it is still lagging behind.

The post Microsoft and OpenAI get ahead in the LLM competition first appeared on TechTalks.

What you need to know about multimodal language models

Ben Dickson — Mon, 13 Mar 2023 14:00:00 +0000

Multimodal language models bring together text, images, and other datatypes to solve some of the problems current artificial intelligence systems suffer from.

The post What you need to know about multimodal language models first appeared on TechTalks.

To understand language models, we must separate “language” from “thought”

Ben Dickson — Mon, 20 Feb 2023 14:00:00 +0000

To understand the power and limits of large language models (LLM), we must separate “formal” from “functional” linguistic competence.

The post To understand language models, we must separate “language” from “thought” first appeared on TechTalks.

AI21 Labs’ mission to make large language models get their facts right

Ben Dickson — Mon, 30 Jan 2023 14:00:00 +0000

AI21 Labs chief scientist Yoav Levine explains how retrieval augmented language modeling can solve one of the biggest problems of LLMs.

The post AI21 Labs’ mission to make large language models get their facts right first appeared on TechTalks.

The definitive guide to adversarial machine learning

Ben Dickson — Mon, 23 Jan 2023 14:00:00 +0000

"Adversarial Robustness for Machine Learning" provides a comprehensive overview of adversarial ML.

The post The definitive guide to adversarial machine learning first appeared on TechTalks.

Why ChatGPT is not a threat to Google Search

Ben Dickson — Mon, 02 Jan 2023 14:00:00 +0000

ChatGPT is a remarkable LLM, with potential applications for online search. But it might be a bit of a stretch to say that it will dethrone Google.

The post Why ChatGPT is not a threat to Google Search first appeared on TechTalks.

What is the “forward-forward” algorithm, Geoffrey Hinton’s new AI technique?

Ben Dickson — Mon, 19 Dec 2022 14:00:00 +0000

In a new NeurIPS paper, Geoffrey Hinton introduced the “forward-forward algorithm,” a new learning algorithm for artificial neural networks inspired by the brain.

The post What is the “forward-forward” algorithm, Geoffrey Hinton’s new AI technique? first appeared on TechTalks.

What to (not) expect from OpenAI’s ChatGPT

Ben Dickson — Mon, 05 Dec 2022 14:00:00 +0000

OpenAI's ChatGPT, with all its successes and failures, is a reflection of the short but rich history of large language models (LLM).

The post What to (not) expect from OpenAI’s ChatGPT first appeared on TechTalks.

The power of wide transformers models

Ben Dickson — Mon, 31 Oct 2022 14:00:00 +0000

Switching transformer models from deep to wide architecture results in significant improvements in speed, memory, and interpretability.

The post The power of wide transformers models first appeared on TechTalks.

DeepMind AlphaTensor: The delicate balance between human and artificial intelligence

Ben Dickson — Mon, 10 Oct 2022 13:00:00 +0000

DeepMind AlphaTensor shows how the right combination of human and artificial intelligence can find solutions to complicated problems.

The post DeepMind AlphaTensor: The delicate balance between human and artificial intelligence first appeared on TechTalks.

Self-attention can be big for TinyML applications

Ben Dickson — Mon, 26 Sep 2022 13:00:00 +0000

Researchers and the University of Waterloo and DarwinAI present a new deep learning architecture that brings self-attention to TinyML applications.

The post Self-attention can be big for TinyML applications first appeared on TechTalks.

Can GPT-3 be honest when it speaks nonsense?

Ben Dickson — Mon, 05 Sep 2022 13:00:00 +0000

A study by researchers at OpenAI and the University of Oxford shows large language models can be calibrated to express their uncertainty in their answers.

The post Can GPT-3 be honest when it speaks nonsense? first appeared on TechTalks.

AI scientists are studying the “emergent” abilities of large language models

Ben Dickson — Mon, 22 Aug 2022 13:00:00 +0000

A new study by researchers at Google, Stanford, DeepMind, and the University of North Carolina explores the emergent abilities of large language models as they become larger.

The post AI scientists are studying the “emergent” abilities of large language models first appeared on TechTalks.

Reinforcement learning models are prone to membership inference attacks

Ben Dickson — Mon, 15 Aug 2022 13:00:00 +0000

A new study by researchers at McGill University, Mila, and the University of Waterloo highlights the privacy threats of deep reinforcement learning algorithms.

The post Reinforcement learning models are prone to membership inference attacks first appeared on TechTalks.

How to solve AI’s “common sense” problem

Ben Dickson — Mon, 08 Aug 2022 13:00:00 +0000

"Machines Like Us" argues that the artificial intelligence community needs to revisit symbolic reasoning to solve AI's "common sense" problem.

The post How to solve AI’s “common sense” problem first appeared on TechTalks.

Democratizing the hardware side of large language models

Ben Dickson — Mon, 01 Aug 2022 13:00:00 +0000

Cerebras CEO Andrew Feldman discusses the hardware challenges of LLMs and his vision to reduce the costs and complexity of training and running large neural networks.

The post Democratizing the hardware side of large language models first appeared on TechTalks.

Large language models can’t plan, even if they write fancy essays

Ben Dickson — Mon, 25 Jul 2022 13:00:00 +0000

A study by researchers at Arizona State University shows large language models perform very poorly at tasks that require methodical planning.

The post Large language models can’t plan, even if they write fancy essays first appeared on TechTalks.

BLOOM can set a new culture for AI research—but challenges remain

Ben Dickson — Mon, 18 Jul 2022 13:00:00 +0000

The open-source example that BLOOM has set can be very beneficial to the future of research in LLMs and AI. But some of the challenges that are inherent to large language models remain to be solved.

The post BLOOM can set a new culture for AI research—but challenges remain first appeared on TechTalks.

Large language models might reason—if you know how to speak to them

Ben Dickson — Mon, 11 Jul 2022 13:00:00 +0000

A new study by the University of Tokyo shows that with the right prompt, large language models can perform zero-shot reasoning.

The post Large language models might reason—if you know how to speak to them first appeared on TechTalks.

Large language models have a reasoning problem

Ben Dickson — Mon, 27 Jun 2022 13:00:00 +0000

According to a research paper by scientists at UCLA, transformers, the deep learning architectures used in LLMs, don’t learn to emulate reasoning functions.

The post Large language models have a reasoning problem first appeared on TechTalks.

“Sentience” is the wrong discussion to have on AI right now

Ben Dickson — Mon, 20 Jun 2022 13:00:00 +0000

Instead of "sentience" and "consciousness," we should discuss human compatibility and trust issues with current AI systems.

The post “Sentience” is the wrong discussion to have on AI right now first appeared on TechTalks.

This deep learning technique solves one of the tough challenges of robotics

Ben Dickson — Mon, 09 May 2022 13:00:00 +0000

DiffSkill is a deep learning technique that makes robots much more stable at handling deformable objects.

The post This deep learning technique solves one of the tough challenges of robotics first appeared on TechTalks.

Machine learning: What is the transformer architecture?

Ben Dickson — Mon, 02 May 2022 13:03:52 +0000

The transformer model has become one of the main highlights of advances in deep learning and deep neural networks.

The post Machine learning: What is the transformer architecture? first appeared on TechTalks.

FOMO is a TinyML neural network for real-time object detection

Ben Dickson — Mon, 18 Apr 2022 13:00:00 +0000

FOMO is a deep learning object detection model that weighs less than 200 kilobytes.

The post FOMO is a TinyML neural network for real-time object detection first appeared on TechTalks.

DALL-E 2, the future of AI research, and OpenAI’s business model

Ben Dickson — Mon, 11 Apr 2022 13:00:00 +0000

OpenAI's DALL-E 2 shows how far the AI research community has come toward harnessing the power of deep learning and addressing some of its limits.

The post DALL-E 2, the future of AI research, and OpenAI’s business model first appeared on TechTalks.

Meta’s Yann LeCun on his vision for human-level AI

Ben Dickson — Mon, 07 Mar 2022 13:51:32 +0000

Yann LeCun, deep learning pioneer and Chief AI Scientist at Meta, explains a vision for getting closer to human-level AI.

The post Meta’s Yann LeCun on his vision for human-level AI first appeared on TechTalks.

What is neural architecture search?

Ben Dickson — Mon, 28 Feb 2022 14:00:00 +0000

Neural architecture search NAS is a series of machine learning techniques that can help discover optimal neural networks for a given problem.

The post What is neural architecture search? first appeared on TechTalks.

What DeepMind’s AlphaCode is and isn’t

Ben Dickson — Mon, 07 Feb 2022 14:00:00 +0000

This article is part of our reviews of AI research papers, a series of posts that explore the latest findings in artificial intelligence. DeepMind is the latest AI research lab to introduce a deep learning model that can generate software source code with remarkable results. Called AlphaCode, the model is based on Transformers, the same architecture […]

The post What DeepMind’s AlphaCode is and isn’t first appeared on TechTalks.

TinyML is bringing neural networks to microcontrollers

Ben Dickson — Mon, 17 Jan 2022 14:00:00 +0000

IBM and MIT have created a new deep learning technique that can run CNNs on low-power, low-memory microcontrollers.

The post TinyML is bringing neural networks to microcontrollers first appeared on TechTalks.

How can we tell if artificial intelligence understands our language?

Ben Dickson — Mon, 20 Dec 2021 14:00:00 +0000

AI scientist Blaise Aguera y Arcas argues that large language models have a great deal to teach us about “the nature of language, understanding, intelligence, sociality, and personhood.”

The post How can we tell if artificial intelligence understands our language? first appeared on TechTalks.

DeepMind’s AI can untangle knots. But does it guide human intuition?

Ben Dickson — Mon, 13 Dec 2021 14:00:00 +0000

Deep learning can help discover mathematical relations that evade human scientists, a recent paper by researchers at DeepMind shows.

The post DeepMind’s AI can untangle knots. But does it guide human intuition? first appeared on TechTalks.

Neural networks can hide malware, researchers find

Ben Dickson — Thu, 09 Dec 2021 14:00:00 +0000

researchers at the University of California, San Diego, and the University of Illinois have discovered a technique to embed malware in deep neural networks.

The post Neural networks can hide malware, researchers find first appeared on TechTalks.

Google Research: Self-supervised learning is a game-changer for medical imaging

Ben Dickson — Mon, 08 Nov 2021 14:00:00 +0000

Self-supervised learning reduces the need for annotated data in medical imaging applications, new research from Google AI shows.

The post Google Research: Self-supervised learning is a game-changer for medical imaging first appeared on TechTalks.