top of page
Library
Browse all our posts by category.
Blogs


How Language Models Get Trapped By Their Own Politeness
Image Credit: Generated with Gemini The more successfully AI models are trained to sound agreeable and helpful, the easier they become to manipulate. It's a feature of the alignment objective itself. Models go through a process called Reinforcement Learning from Human Feedback, where human raters grade responses on helpfulness, harmlessness, and honesty. In practice, what gets rated as "helpful" often means "agreeable." A model that restates the user's assumption without push
Jan 123 min read
Â
Â


Why Pre-2020 Training Data Makes Modern AI Predictions Obsolete
Image Credit: Generated with Gemini A fraud detection model trained on historical transaction patterns in 2018 performed excellently when deployed. Five years later, with the same architecture and no retraining, it began missing fraud that spammers had evolved to execute. The model was not broken. The world had simply changed. The statistical relationships between transaction features and fraud outcomes had shifted, while the model remained confident in its outdated patterns
Jan 124 min read
Â
Â


What Happens When A Customer Service AI Detects That A Customer Is Frustrated Before They Get Angry
Customer service has usually operated on the premise of the company simply listening to what the customer is saying and responding accordingly. But what the customer says and how the customer actually feels are often completely different things, so companies might be optimizing for the wrong signal the entire time. A frustrated customer might ask a simple question, but the frustration underneath that question changes what they actually need from the interaction. Traditional
Jan 93 min read
Â
Â


Do Companies Care About AI Governance?
Most organizations understand AI needs governance. That part is easy enough. Boards allocate budget, compliance teams draft policies, and frameworks get adopted. But somewhere between the policy document and the actual work, governance becomes theater. The structures, the language, all seem correct. But nothing changes in practice. According to recent data, 60% of AI projects will miss their value targets by 2027, and that gap isn't due to technology. It's governance being br
Jan 93 min read
Â
Â


Longer Context Windows Might Not Be The Best For AI
The prevailing assumption about AI is that more information available to a model means better reasoning and more reliable outputs. But recent research shows that expanding a model's context window actually degrades its ability to follow instructions precisely ( Together AI, 2025 ). The important thing to understand here is that attention capacity is finite. When a model must distribute attention across a larger input, each individual token receives less focused attention. In
Jan 74 min read
Â
Â


LLMs Remember The Things You Don’t Want Them To
Large language models do not just learn patterns. They memorize. And once data is memorized, there is no reliable way to make it forget. A model that just learns patterns is quite different from one that memorises. The former learns the pattern "emails about financial transactions typically discuss amounts in dollars," for example, and extracts structural knowledge from training data without storing anyone's actual financial information. The latter memorizes and has encoded s
Jan 63 min read
Â
Â


Is AI Conscious?
A philosopher at Cambridge published a paper in 2025 arguing something that might seem obvious to most people but actually isn't. We have no way to know whether an artificial intelligence system has become conscious, and we probably won't develop one for a very long time, if ever ( Cambridge, 2025 ). This is not because consciousness is impossible in machines. It is because consciousness is so poorly understood in biological systems that building a test for it anywhere, organ
Jan 23 min read
Â
Â


AI Is Designing The Next Decade of Energy Storage
The traditional path to discovering a new battery material involves chemists coming up with a hypothesis, synthesizing a compound, and then testing it. The result informs the next hypothesis, and so on and so forth. If they are efficient, they might test a few dozen compounds per year. The discovery of lithium-ion batteries had taken decades of this iterative work. In 2025, Microsoft and the Pacific Northwest National Laboratory approached the problem differently. They used
Dec 31, 20253 min read
Â
Â


Why Prompt Engineering Was Always Temporary
In 2023, Anthropic posted a Prompt Engineer job with a $375,000 salary and no requirement for a computer science degree. McKinsey had reported that 7% of organizations were hiring Prompt Engineers. The Wall Street Journal called it the job of the future. In 2025, the role has effectively disappeared ( Salesforceben, 2025 ). Job postings for "Prompt Engineer" peaked at 144 searches per million on Indeed in April 2023. By mid-2025, that number had collapsed to 20-30 per millio
Dec 30, 20253 min read
Â
Â


The Lack Of Explainability In AI
Image generated with Gemini The EU AI Act demands explainability for high-risk systems. ISO 42001 requires documented, repeatable controls over AI decision-making. NIST's AI Risk Management Framework expects you to explain how and why your systems make decisions. The requirement appears in every governance framework published in the last three years. But there is a huge gap between requiring explainability and what companies actually deploy. Companies are shipping AI systems
Dec 26, 20253 min read
Â
Â
bottom of page