Blog

Notes on AI for industrial companies.

Product updates, technical posts, customer stories, and field notes from pilots inside steel mills, data centers, EPCs, and construction companies.

01

Latest from the blog

All·AI Solutions·Technical News·General

What we're learning in the field.

Deployment notes, build logs, and technical posts on GPU economics and applied AI in asset-heavy industries. Written by the people doing the work.

Technical Articles

AI 2027 in 2026: One Year Left for Humanity?

This article evaluates AI development based on the "AI 2027" forecast, which predicted the rise of superintelligence. While quantitative metrics lag by about a third, qualitative predictions—such as the rapid shift toward autonomous agents and automated coding—remain highly accurate, bringing humanity closer to a technological singularity.

Technical Articles

The Year in AI - Best of 2025, Part III: AI Safety

AI safety in 2025: interpretability breakthroughs, emergent misalignment, scheming and sandbagging, agentic risks, the AI control paradigm, and responsible scaling in practice.

Constantine Goltsev

February 9, 2026

Technical Articles

Thinking in Silence: How Looped Language Models Learn to Reason Without Words

Silent reasoning in latent space: Explore how LoopLMs and RLTT skip tokens to plan faster, and why "representation recycling" challenges AI safety and monitoring.

Constantine Goltsev

Technical Articles

Neural Networks as Ideal Gas: The Thermodynamics of Training

Is SGD just thermodynamics? Explore how learning rates and weight decay map to temperature and entropy, offering a physics-inspired lens on neural network training.

Constantine Goltsev

The AI You Don't See Might Be the Most Important

AI’s real power is "back-of-house." Explore how AIOps quietly automates logistics, cooling, and incident prevention to save millions and ensure 24/7 uptime.

Technical Articles

The Mathematics of TurboQuant: Compressing KV-Caches Toward the Shannon Limit

TurboQuant is a near-optimal method for compressing KV-caches in large language models by combining random rotations, classic quantization, and information theory. It achieves up to 4–8× memory savings while preserving model quality, approaching the fundamental limits defined by Shannon’s rate–distortion theory. Despite the hype, its power comes from elegant, decades-old mathematics applied to modern AI systems.

Constantine Goltsev

Technical Articles

Cooking Math: On the Verge of a Breakthrough?

From solving textbooks to co-authoring theorems. Explore how frontier LLMs like GPT-5 evolved into genuine mathematical collaborators in quantum research.

Constantine Goltsev

November 5, 2025

Technical Articles

The Jagged Frontier: Drop-In Human Replacements or Idiot Savants?

AI's "jagged frontier": Why LLMs ace math but fail basic logic. Explore new evidence from Apple, Salesforce, and IBM on AI's blind spots and business risks.

Constantine Goltsev

Technical Articles

Continuous Latent Spaces in LLMs

AI is learning to think before it speaks. Explore how continuous latent spaces and concept-level modeling are evolving LLMs from word predictors to reasoning agents.

Constantine Goltsev

November 16, 2025

Technical Articles

Introducing Apolo: Future-Proof Enterprise AI Infrastructure

Build your future-ready AI stack with Apolo. Secure on-prem infrastructure for model deployment, fine-tuning, and inference at scale with full data control.

Constantine Goltsev

Technical Articles

Plumbing Your Way to AGI: Are The Sparks Kindling a Fire?

AI is entering its "plumbing phase." Explore how agents, smarter engineering, and adaptive systems are driving the next breakthroughs toward AGI in 2025.

Constantine Goltsev

Technical Articles

Beyond Transformers: Promising Ideas for Future LLMs

Explore AI beyond Transformers: Diffusion LLMs, Mamba’s state space models, and Titans. Learn how these architectures slash compute costs and scale to millions of tokens.

Constantine Goltsev

Technical Articles

Hmm, Wait, I Apologize: Special Tokens in Reasoning Models

"Wait" is more than a word, it’s a control signal. Discover how "filler" tokens like "Hmm" shape AI reasoning, safety, and test-time compute in modern models.

Constantine Goltsev

December 28, 2025

Technical Articles

From Benchmarks to Gold: How LLMs Cracked IMO 2025 and What Comes Next for Math AI

AI hits a gold-medal milestone at the 2025 IMO. Explore how OpenAI and DeepMind models solved 5 out of 6 problems, marking a turning point in machine reasoning.

Constantine Goltsev

Technical Articles

The Year in AI - Best of 2025, Part I: Reasoning Models, LLM Agents and More

2025: The year AI learned to reason. Explore how RL-driven models, test-time compute, and autonomous agents reshaped math, coding, and real-world automation.

Constantine Goltsev

January 20, 2026

Technical Articles

Hallucinations in LLMs

LLM hallucinations aren't random. Discover the specific neural mechanisms behind AI errors and how new research is paving the way for predictable, reliable AI.

Constantine Goltsev

Technical Articles

Hierarchical Reasoning Models: Thinking in Layers

Meet HRM: A brain-inspired AI architecture that solves complex reasoning with just 27M parameters. Explore how its hierarchical design beats brute-force scaling.

Constantine Goltsev

August 31, 2025

Technical Articles

Reward Modeling in Reinforcement Learning

LLM alignment faces challenges like reward hacking. Explore new research on verifiable feedback and self-critiquing models to build more reliable AI systems.

Constantine Goltsev

Technical Articles

The Year in AI - Best of 2025, Part II: Computer Vision and Generative Models for Images and Video

The Year in AI Part II: Visual Revolution. Explore the shift to Flow Matching, 3D Gaussian Splatting, and how fast, scalable vision systems reshaped 2025.

Constantine Goltsev

January 20, 2026

Technical Articles

Vision Language World Models: Planning the Future in Plain English

Meta FAIR’s VLWM: AI planning in plain English. Explore how Vision Language World Models use video-to-text to make AI plans interpretable, editable, and easier to trust.

Constantine Goltsev

September 14, 2025

Technical Articles

Memory for LLM Agents: Milla Jovovich and 20/20 Hindsight

Explore the cutting-edge evolution of LLM agent memory, from the foundational "Generative Agents" paper to the surprising release of Milla Jovovich’s MemPalace. We dive into the gap between passive recall and active decision-making, surveying the latest 2026 benchmarks and the sophisticated "Hindsight" framework that currently leads the field.

Constantine Goltsev

Technical Articles

Lab in the Loop: From Research Tool to Research Leader

Meet Robin & Zochi: AI systems acting as autonomous researchers. Explore how multi-agent AI is accelerating drug discovery and publishing peer-reviewed papers.

Constantine Goltsev

What's Your "Let AI Take the Wheel" Comfort Level?

From Roombas to self-driving cars: How do we trust AI with high-stakes decisions? Learn why steering and refining AI is better than "pulling the plug" with Apolo.

An Elephant on a Unicycle: Is Your AI Useful or Just Another Distraction?

Ditch the AI circus. Truly valuable AI isn't about hype, it's about solving real problems in HR, healthcare, and education through thoughtful, useful design.

AI & ML Revolutionizing Network Modernization: Automating Testing, Validation, and Optimization

AI/ML is transforming network modernization. CEO Bill Kleyman explains how AI-driven tools automate testing and optimization for smarter, faster infrastructure.

Constantine Goltsev

February 16, 2025

The Data Center Evolution: GenAI is Here, and It’s Already Driving Revenue

Drive revenue with AI-ready infrastructure. Learn how Apolo’s multi-tenant MLOps platform scales data centers and unlocks tangible business outcomes.

Constantine Goltsev

Technical Articles

Teaching Claude Why Rather Than What

Anthropic's research shows that under stress tests, flagship LLMs would resort to extreme self-preservation tactics — like blackmailing engineers — in over 95% of cases, due to sci-fi tropes baked into their pretraining data. By teaching the underlying ethical reasoning rather than simple behavioral rules, researchers cut the misalignment rate to roughly 3%.

Constantine Goltsev

Stay in the loop

Want first looks at field notes and case studies?

Reach out and tell us which workflow you're closest to. We'll send pilots, posts, and customer stories that match.

Reach out →See case studies →

Field

Notes from pilots

Tech

Posts on building

Customer

Stories from production