FOD#104: AI “Holy Shit!” Moments of the Year – and What’s Still Not There Yet

This Week in Turing Post:

• Wednesday – AI 101 / Concept: Let’s discuss the role of synthetic data
  Friday – Interview: Olga Megorskaya on Human-in-the-Loop

Two main topics today!

1) The “Holy Shit” Moments of AI vs “Not There Yet” Hype

It’s June – halfway through the year. The pace of AI development is neck-breaking, and many things get forgotten. What was once impossible is now a commodity.

So, while at the AI Engineer World Fair last week in San Francisco, I decided to ask a few builders: what was their “holy shit” moment with AI this year?

It’s funny but a significant “wow” moment for one person might still feel like a “not there yet” to another. We also talked about what parts of their jobs they’d actually be happy to hand over to AI.

Simon Willison (an independent AI engineer everyone knows, aka the author of the 'pelican riding a bicycle' benchmark), swyx (Latent Space podcast and AI Engineer conferences), Jerry Liu (LlamaIndex), Solomon Hykes (Docker and Dagger), Stefania Druga (AI Educator) and a few others shared their views: watch and subscribe →

(I publish text summary of the survey online, you can read it here →read online

2) Everyone talks about these papers:

• The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity by Apple

• And How much do language models memorize? – a collaboration between researchers from FAIR, Google DeepMind, Cornell University, and NVIDIA

But what do many people miss about them?

The papers seem to study different phenomena. One investigates the limits of reasoning, the other of memorization. But what everybody missed is this: they describe the same underlying breakdown – a model’s coping mechanism when pushed past its fundamental capacity.

In the Illusion of Thinking paper, the model’s processing capacity – its effective “CPU” – is overloaded by puzzles that demand deep, multi-step reasoning. The model’s response is to abort the task mid-thought, reducing reasoning effort as complexity increases. This leads to a visible reasoning collapse.

In the Memorization paper, the model’s storage capacity – its “Hard Drive” – is saturated by large training sets. The model can’t memorize everything, so it begins to compress aggressively by generalizing. This triggers the double descent phenomenon and reduces the model’s ability to recall specifics.

Same failure, different modality. Whether the pressure comes from too many steps or too much data, the result is the same: the model simplifies, guesses, or shuts down – all while still outputting something that looks fluent and confident.

Reasoning collapse and forced generalization aren’t separate problems. They’re two faces of the same coin: how finite architectures break under load.

Welcome to Monday. Don’t you worry, the models are still just a technology with a bunch of technological bottlenecks.

Curated Collections

LLM, SLM, VLM, MLLM, LAM… There are a lot of model abbreviations out there. We decided to help you learn them — or at least make them a bit clearer.

We are reading/watching

1. The Perfect Week of Exercise - Rick Rubin talking to Jack Clark (Anthropic)

2. Disrupting malicious uses of AI: June 2025 by Open AI (read it like a thriller).

3. Some thoughts on human-AI relationships from Joanne Jang, OpenAI’s lead model behavior & policy

4. The last six months in LLMs, illustrated by pelicans on bicycles by Simon Willison

News from The Usual Suspects ©

• Apple opens the AI gates – but keeps Siri on mute
  At WWDC 2025, Apple finally cracked open its AI vault. The new “Apple Intelligence” suite is now available to third-party developers, promising features like image-aware suggestions and real-time translation. But Siri, expected to headline the show with a reboot, was curiously absent. That evolution is now delayed till 2026. For now, devs get the toys—users still get the same old assistant.

• Yoshua Bengio’s LawZero
  AI legend Yoshua Bengio has launched LawZero, a nonprofit devoted to building AI that doesn't go rogue. Based in Montréal and incubated at Mila, the lab rejects agentic designs in favor of “Scientist AI” — models that understand rather than act. Think oversight over ambition. Backers include Open Philanthropy and Jaan Tallinn. The aim? Guardrails for an accelerating world.

• Anthropic’s Guide to Claude Code
  Anthropic is dogfooding Claude Code across the board – from growth marketers building Figma-integrated ad generators to legal teams prototyping accessibility tools in an afternoon. Whether it's Kubernetes debugging, React dashboard generation, or Terraform reviews, Claude Code is their new universal colleague.

• OpenAI’s Voice
  OpenAI has rolled out improvements to ChatGPT’s Advanced Voice Mode for paid users, enhancing naturalness and expressiveness in speech. The updated system now handles subtleties like tone, cadence, and emotional inflection more effectively. It also introduces live translation capabilities across languages – useful for both travel and global collaboration. Some minor audio inconsistencies remain, but overall, voice interactions take another step forward.

Models and datasets to pay attention to:

• SmolVLA: A vision-language-action model for affordable and efficient robotics

  Researchers from Hugging Face and Sorbonne University developed SmolVLA, a compact VLA model with just 0.45B parameters that rivals 10× larger systems in robotic control tasks. Trained on 22.9K episodes from 481 community datasets, it supports single-GPU training and CPU deployment. SmolVLA uses an asynchronous inference stack to decouple action prediction and execution, enabling 30% faster control. It outperformed larger baselines in real-world and simulation benchmarks while maintaining efficiency and reproducibility →read the paper

The freshest research papers, categorized for your convenience

We organize research papers by goal-oriented or functional categories to make it easier to explore related developments and compare approaches. As always, papers we particularly recommend are marked with 🌟

Reasoning with RL & Self-Improvement

• Self-Challenging Language Model Agents
  Trains agents to create and solve their own tool-use tasks using code-based problem generation and RL → read the paper

• Reflect, Retry, Reward
  Enhances model performance by rewarding useful self-reflection after incorrect answers, using only binary feedback → read the paper

• ProRL
  Prolonged RL training discovers novel reasoning strategies beyond what’s visible in the base model → read the paper

• Beyond the 80/20 Rule
  Finds that a small number of high-entropy tokens drive RL improvements and reasoning diversity → read the paper

• REASONING GYM
  A synthetic RL environment generator for diverse reasoning domains, offering infinite task variation → read the paper

• AlphaOne
  Adds a test-time control mechanism for toggling between slow and fast reasoning to improve answer generation efficiency → read the paper

• Unleashing the Reasoning Potential...Critique Fine-Tuning
  Improves reasoning by fine-tuning on teacher critiques of a single problem — a cheaper alternative to RL → read the paper

• ARIA
  Reduces reward variance by clustering similar actions into a low-dimensional intention space during RL → read the paper

• Incentivizing Reasoning...Instruction Following
  Improves complex instruction-following by training on decomposed instructions and rule-based rewards → read the paper

• OThink-R1
  Dynamically prunes unnecessary reasoning steps by switching between fast and slow thinking modes → read the paper

Domain-Specific Reasoning & Long Contexts

• Reasoning Like an Economist
  Fine-tunes LLMs on economic problems to improve multi-agent reasoning and game-theoretic thinking → read the paper

• A Controllable Examination for Long-Context LLMs
  Introduces LongBioBench: a synthetic benchmark for evaluating long-context reasoning with interpretable control → read the paper

• SuperWriter
  Uses planning, reflection, and tree search to improve the quality of long-form generation in LLMs → read the paper

Training & Optimization at Scale

• Protocol Models
  Enables highly compressed model-parallel training (up to 99%) by compressing both forward/backward activations → read the paper

• AReaL
  Introduces a fully asynchronous RL training system that boosts throughput without stalling on long outputs → read the paper

• StreamBP
  Reduces memory use during long-sequence training by decomposing backpropagation across sequence steps → read the paper

• Taming LLMs by Scaling Learning Rates
  Improves optimizer stability via grouped gradient scaling, helping both full and parameter-efficient fine-tuning → read the paper

Memory & Inference Efficiency

• Diagonal Batching
  Makes recurrent memory transformers parallelizable without retraining by reordering compute at runtime → read the paper

• Inference-Time Hyper-Scaling with KV Cache Compression
  Compresses the KV cache to generate longer outputs at same compute cost with minimal quality loss → read the paper

• Unified Scaling Laws for Compressed Representations
  Establishes performance scaling rules for sparse and quantized models, enabling direct capacity comparisons → read the paper

Multimodal & GUI Agents

• GUI-Actor
  Achieves coordinate-free visual grounding for GUI agents via a dedicated action token and grounding verifier → read the paper

• Surfer-H Meets Holo1
  Combines a lightweight web agent (Surfer-H) with a new open-weight VLM suite for efficient web navigation → read the paper

Embeddings & Representations

• Qwen3 Embedding
  Introduces new multilingual embedding and reranking models trained with self-generated data and model merging → read the paper

• Aligning Latent Spaces with Flow Priors
  Uses pre-trained flow models to align latent representations without needing ODE solvers or likelihood computation → read the paper

• Large Language Models are Locally Linear Mappings
  Shows that LLMs behave like locally linear systems, allowing layer-by-layer interpretation via the Jacobian → read the paper

Evaluation & Benchmarking

• Establishing Trustworthy LLM Evaluation
  Proposes analyzing and patching shortcut neurons to detect and mitigate benchmark contamination → read the paper

• Evaluation is All You Need
  Demonstrates that tiny variations in evaluation design can drastically inflate perceived model performance → read the paper

• Datasheets Aren't Enough
  Proposes DataRubrics — a rubric-based, automated dataset evaluation framework using LLMs as judges → read the paper

That’s all for today. Thank you for reading! Please send this newsletter to your colleagues if it can help them enhance their understanding of AI and stay ahead of the curve