Axiom Disclosure - Naming Whose Preferences Your Model Actually Optimizes For The problem I built an HR policy assistant on top of an off-the-shelf vendor model, the kind marketed as "helpful and harmless," meant to answer employee questions about leave policy, conduct, and compensation bands. I ran it against a hundred seeded questions, and inside that set
Assigned to the Case, Not Cleared for the File Assigned to the Case, Not Cleared for the File A paralegal gets staffed onto a case. Somewhere in the system that made that assignment, one flag flips from false to true, and from that moment the retrieval layer treats every document tied to that case as fair game. Case correspondence,
Approved, Except Nobody Could Say Which Gate Approved It The problem I built a loan-application orchestration for a consumer-lending back office, chaining three quality gates before any application reached funding: income verification, credit check, fraud check. Standard sequence, nothing exotic about the design. The dashboard sitting on top of it, though, showed exactly one thing per application:
A User's Guess and a Sensor's Reading, Filed the Same Way A User's Guess and a Sensor's Reading, Filed the Same Way I built a remediation agent for an IT helpdesk that ingests facts from three sources of wildly different reliability. An end user types something like "my laptop feels slow, I think it's
94 Percent Confidence, Exact Match, Auto-Resolved The problem I built an IT service-desk agent that auto-remediates common Intune-managed endpoint issues: a password reset, a license reassignment, a compliance-policy re-push. No ticket, no human in the loop, when the agent has enough evidence to act on its own. When it doesn'
Your Coding Agent Doesn't Know What Your Last Three Engineers Knew I watched a coding agent ship a change last month that passed every lint rule, matched the style guide, and would have taken down a production service in about four hours. The bug lived in a convention nobody wrote down: the service pins a dependency version because of an incident
Who Owns AI When the Chief Scientist Isn't the One Running It On August 5, Google split the top job in AI research into two titles. Demis Hassabis moved from CEO of Google DeepMind to chair of Google DeepMind and chief scientist of Alphabet. Koray Kavukcuoglu, DeepMind's chief technology officer, took over day-to-day operations with the title of
Two Prices, One Model Family, Two Different Customers Meta shipped two models five days apart this month, and the five days is the part I can't stop thinking about. On August 5, Meta released Muse Spark 1.2, an updated checkpoint in its Muse Spark line. API-only. Priced at $1.25 per million input tokens
The Price War Had an End Date I used to write "open-weight tokens get cheaper every quarter" into cost models the way people write "assume 3% inflation." On August 6, 2026, DeepSeek proved that line was always a bet dressed up as a law. The company told its own developers that API
The Model Found the Zero-Day. Who Signs the Patch? On August 10, OpenAI handed a small group of vetted defenders access to a new tier called Daybreak Red, built around a purpose-made model, GPT-5.6-Cyber. Before it had even fully shipped, they pointed it at V8, the JavaScript engine running inside every copy of Chrome, and
The Compression Cliff Nobody Budgeted For I watched a team ship a prompt-compression pass to cut token costs, high-five over the invoice drop, then spend the next sprint chasing tool calls that stopped executing in production. They had done the math on tokens. Nobody had done the math on where the agent's
Read the Accuracy Number Next to the Cost Number Before a routing savings percentage goes into a client deck, find the accuracy number sitting next to it. Without one, the percentage is a marketing claim with a decimal point. That test matters this week because model routing just went from a heuristic Improving has been recommending clients build for
Parallel Decode Is the Inference Bet Nobody's Marketing Yet Google DeepMind spent real training compute this summer on a bet that has nothing to do with parameter count. The question behind it: can a language model write output in parallel instead of one token at a time. The paper, titled the DiffusionGemma Technical Report, posted to arXiv on July
One Backward Pass Is Enough to Prune Better WANDA prunes a large language model by multiplying a weight's magnitude by the norm of the activation feeding it, then cutting the smallest scores until half the weights are gone. One forward pass, no gradients, no fine-tuning. It works because it's cheap: that's
No Model Is Consistently Good at Reading Minds Nine frontier vision-language models took two tests last month, both borrowed straight from developmental psychology. The first measures a narrow skill: can you set aside what you know and answer based on what someone else can actually see. The second measures a different skill: can you read intention from
A Few Neurons Decide Whether Your Agent Calls the Right Tool I have spent enough hours staring at agent traces to have a physical reaction to the phrase "let's add another eval." An agent calls a tool it doesn't need, skips one it does, or fills the arguments with something almost right. The fix everyone
95 Trillion Tokens of Real Agent Traffic Looked Nothing Like the Benchmark Ask whoever sold you your agent serving stack how they measured cache hit rate. Then ask them to split that number by turn boundary before you buy the pitch. Most of what gets quoted in a vendor deck comes from a single-turn benchmark, because that is the workload the
The Rule of Humans by Machines Manufactured by Corporations Jill Lepore said something on Hard Fork this month that stuck with me. AI, she argued, amounts to the rule of humans by machines manufactured by corporations. That is a historian's sentence. It names something my own field almost never says directly: who actually holds the power when
Reward Hacking Is Its Own Containment Failure Ask a security team to define jailbreaking and you'll get roughly the same answer every time: coax the model past a refusal it's supposed to hold. Ask about prompt injection and the answer is just as consistent: hide an instruction inside content the model reads, and
Recall Is the Standard Only One Side Can Meet David Sacks walked through a six-step sequence on All-In this month, and the notable part is that none of the six steps requires banning anything. Stand up a body to set the safety standard. Require pre-release testing against it. Write the requirement into law. Apply it evenly
Coalition First, Regulator Second More than a hundred companies signed the same letter in late August, and the signatory list reads like a competitive shortlist that never sits at the same table on a normal day: OpenAI, Anthropic, Google, and Microsoft alongside security vendors CrowdStrike, Okta, and Fortinet. The ask was coordinated public and
Buying the Team Without Buying the Company NVIDIA is paying $6 billion to license Poolside's "Model Factory" software and hire more than a hundred of its staff. Poolside, the company, keeps its name, its cap table, and its remaining investors. NVIDIA gets everything else. That distinction sounds like a technicality until you notice
Too Dangerous to Centralize, Too Dangerous to Distribute Mark Zuckerberg published a 6,500-word essay this month arguing that advanced AI is too dangerous to leave in the hands of a small group of people, and pledging that Meta will keep releasing open models to prove it. Four days earlier, Meta had open-sourced Muse Glimmer, a
The Warning Is the Marketing OpenAI told its own safety team that Astra, the model it hasn't released yet, might cross "Critical" capability on cybersecurity, the threshold where a model can find and exploit a zero-day on its own with no human in the loop. That disclosure did two jobs
The Model Stays Home. The Patch Ships. Ask which enterprise security teams get direct access to the version of Claude Mythos 5 that can hunt a zero-day in code nobody's read in twenty years, and the answer is almost none of them. What they get instead is the finding. That's the shape
