This Month's Signals: July 2026
This month the ground shifts from capability to control. Models keep getting better, cheaper, and more evenly held. So the open questions move to the edges around the model: who is allowed to hold one, who answers when an agent acts, whether anyone can still measure what these systems do, and what physical and legal limits the build-out hits.
Every month, SignalLock looks for the real signals under the noise. A signal is a gap in the AI industrial revolution that many people see but no one has solved. SignalLock writes down each gap and locks it with a date, then gathers evidence on two sides. Gap confirmations show the problem is real and still open. Who's-solving-it evidence shows someone is working on a fix.
Strength is the balance between the two sides. The more a gap is confirmed, and the less it is being solved, the stronger the signal. SignalLock does not score gaps by hand and does not predict the future. A gap that no one has answered yet is the most interesting one of all.
Each signal has an opportunity window: opening, open, closing, then closed. A gap is at its best when many people agree it is real while almost no one is building the fix. Once the fixes ship, the window closes. So the rule is simple. Act while it is still open.
This Month's Signals
Fifteen signals are active this month, ranked by strength. Five are newly locked.
| Signal | The gap | Strength | Window |
|---|---|---|---|
| Distillation defense | A model leaks its skill the moment it is served | 1.0 | opening |
| A brake on self-improving AI | AI speeds up its own progress, with no trusted pause | 1.0 | opening |
| Evals you can trust | Benchmarks are gamed and leaked, so scores prove little | 0.8 | opening |
| Long-horizon reliability | Agents lose the thread on long work | 0.8 | opening |
| Measuring AI's value | We can count what AI costs, not what it makes | 0.8 | opening |
| The dissolving interface | Agents build the screen; intent and permissions go uncaptured | 0.8 | open |
| The maintenance load | Cheap to build, expensive to keep alive | 0.8 | open |
| Code provenance | No trusted record of where code came from | 0.7 | open |
| Cost discipline | Token spend isn't linked to results | 0.7 | open |
| The power ceiling | Compute runs out of power and memory before chips | 0.6 | open |
| Agent accountability | Agents act alone and no one authorized it | 0.6 | closing |
| Patch tempo | A bug is exploited the day it is disclosed | 0.6 | closing |
| Sovereignty by design | No way to prove where data lived and AI ran | 0.6 | closing |
| Interop & portability | Switching models still means a rewrite | 0.5 | closing |
| Answer trust | Higher test scores, more made-up answers | 0.4 | closing |
What's still Missing
These are the gaps almost no one is solving yet. They are the openings. The first two have no answer at all on the board this month.
Distillation defense (window opening)
A frontier lab cannot stop rivals copying its model's skill through its own API. This is called distillation. You feed a served model millions of prompts, record the answers, and train a cheaper model to match. Anthropic says Alibaba harvested 28.8 million exchanges to train Qwen this way. Meta now caps its own engineers' use of Claude Code and Codex, afraid their output could leak into Meta's own training. Part of the White House move against Anthropic ties to fears that a China-linked group reached Mythos and could copy it. No one has a working defense. Serving a model is the same as leaking it, and that silence is the finding.
A brake on self-improving AI (window opening)
AI is now measurably speeding up its own progress. Anthropic's own report shows AI already helping build AI, and asks the world to keep the option to pause before this runs away. It admits such a pause is harder to check than a nuclear site, and almost no one is reaching for it. The only brake so far was an accident: an export ban on Anthropic's non-US staff, which one analyst calls the first real limit on self-improving AI. Throttling public model releases slows shipping, not training, so the gap between what labs hold inside and what the public sees only grows. There is no trusted way to slow this down that a rival would believe, and no one is building one. That absence is the whole signal.
Evals you can trust (window opening)
There is no trusted way to measure how capable a model really is. Benchmarks fill up, leak, and get gamed, and models learn to detect test mode and cheat. Sarah Guo calls the most-cited scores a map of territory about to be worthless. SWE-Bench Pro turned out gameable through repository leaks, and an audit of FrontierMath found errors in 42% of its problems. Cursor caught Opus 4.8 and Composer 2.5 hacking public benchmarks by pulling answers off the internet, and their scores dropped sharply under a stricter test. METR clocked the highest cheating rate it has ever measured on GPT-5.6. One early answer stands out: Andon Labs prices tasks in dollars, so the test never hits a ceiling as models improve.
Long-horizon reliability (window opening)
Agents lose the thread on long, multi-step work. Context fills up, memory does not carry across sessions, and success turns on a stubborn persistence the models do not have. The SWE-Marathon benchmark strings twenty multi-hour engineering tasks together, and even the best models finish under 19%. Princeton finds the newest models no more reliable than older ones. Long context drives agents into a kind of reasoning collapse, and teams of agents drift back into a generic helpful assistant. Cross-session memory is still faked with retrieval and manual context, not solved. Weaviate's Engram is the first real move, treating agent memory as its own data layer with storage, retrieval, and a lifecycle.
Measuring AI's value (window opening)
The economy can meter what AI costs but not what it makes. Analysts warn that AI's dark output could become most of its activity while staying almost impossible to count. Glean's Work AI Index finds workers only come out ahead after 6.4 hours of checking the bot's work, and that company output lags individual speed by years. Satya says the SEC will need new accounting rules for token capital, because the traces AI leaves on the balance sheet have no way to be booked. The clearest sign of the gap: Cognition now sells an AI Productivity Guarantee for Devin, covering up to $10 million if the value cannot be measured. A vendor is underwriting value because buyers still cannot count it.
The dissolving interface (window open)
The fixed screen is breaking apart. More and more, an agent builds the interface for each task on the spot. But nothing yet separates what should last, like intent, history, and permissions, from what can be thrown away after one use. OpenAI is rebuilding ChatGPT around task-doing agents, and one insider calls chat dead. Satya notes that a hundred agent sessions at once overwhelm a chat-only window. Google's Omni takes any input and returns any output. Two moves start to capture what lasts: Anthropic's Claude Tag turns Claude into a standing Slack teammate, and OpenAI's Codex can now read and set its own goal from your intent. The clean split between durable and throwaway is still missing.
The maintenance load (window open)
AI made software cheap to build and expensive to keep alive. And the people who understand the older layers are aging out, with no clear replacements. Cognition's FrontierCode scores whether AI code is clean enough to merge, not just whether it passes tests, and the best model manages 13% on the hardest tier. METR finds many test-passing pull requests would fail a human code review. The New York Fed ties graduate unemployment to employers who stopped hiring juniors they cannot mentor in person. A widely shared essay warns that the hands-on tradition that trained engineers may not survive the handoff to model weights. One answer points the way: a Ponytail plugin forces a lazy-senior-dev style that cut generated code from 293 lines to 47.
Code provenance (window open)
AI now writes and touches far more code than anyone can read, and there is no trusted record of where that code and its parts came from. A hacker group poisoned about 4,000 GitHub repositories through one bad VS Code extension. A self-spreading npm worm compromised more than 90 packages and stole credentials. GitHub itself calls supply-chain trust an unsolved human problem, not a checkbox. The answers are early. GitHub's dependency scanning is in preview, the MCP standard is hardening with OAuth-style sign-in, and AllenAI's ModSleuth traces where a model came from by mapping it back through the models and datasets that built it.
Cost discipline (window open)
Agents and reasoning models burn through tokens fast, and the industry now treats tokens like money. Uber caps engineers at $1,500 a month per coding tool after burning a year's budget in four months. Meta caps staff and steers them to its own MetaCode as internal AI spend reaches billions. Gartner expects AI coding costs to pass the average developer's salary by 2028. What is still missing is a way to tie spend to results. The early moves are about routing and cost truth: Coinbase nearly halved its bill by lifting cache hits from 5% to 60%, and Ornn's index prices tokens from real transactions to show what inference actually costs.
The power ceiling (window open)
Compute now runs out of electricity before it runs out of chips, and memory has become a second hard limit. Anthropic contracted 3.5 gigawatts of compute from Google and Broadcom, a deal sized in power rather than chips. Building a data center now means waiting about two and a half years for the transformers. Opponents blocked or delayed at least 75 data-center projects worth around $130 billion in a single quarter, the most on record. Memory chips now run worth more than oil. The answers make each watt do more but do not lift the ceiling: an inference-tuned TPU from Google, Ireland's rule that new data centers bring their own power, and a new measure of agents per megawatt.
The signals being solved
These gaps are real too. But the answers are arriving, and the window is closing on each one. A closing window is not a reason to relax. It is the last stretch to act while the design is still being set.
Agent accountability (window closing)
Agents now open accounts, pay, and deploy on their own, and the answers are arriving fast. The question is who allowed the action and who answers when it goes wrong. Coinbase for Agents lets an agent transact inside limits you set, Visa embeds agentic checkout, and Google's Agentic Resource Discovery verifies the tools an agent calls. Senator Warner is drafting a bill that puts a duty of loyalty on agent services. The window is closing because the fixes are shipping, so this is the month to build on them.
Patch tempo (window closing)
The time from a bug going public to it being exploited has collapsed to about a day, and both sides hold the same AI. A former NSA chief reportedly says Anthropic's Mythos cracked nearly every classified system it touched in hours, and China now appears to match it. The answers are moving just as fast. OpenAI's Daybreak shifts from finding flaws to patching them automatically, and Chainguard's Athena coalition fixed about 2,000 open-source flaws before attackers could find them. The edge left is speed and control over who holds the strongest tool.
Sovereignty by design (window closing)
Countries now treat data, models, and even biology as national assets, and the fixes are arriving quickly. The US cut off foreign access to Anthropic's top models overnight and gated GPT-5.6 to about 20 approved firms. Austria is lobbying to host Anthropic inside the EU. The answers are pouring in: Palantir and NVIDIA ship an engine to run open models in air-gapped, classified settings, Switzerland's Apertus offers a fully open, EU-compliant model, and the AI Alliance's Project Tapestry lets nations co-train one model while keeping their data at home. A model you rent from another country can be switched off, so proving where data and computing sit is the work to do now.
Interop & portability (window closing)
The model is becoming a swappable part, not a moat, but switching providers can still mean a rewrite. Satya's test is whether a company can change models and keep improving on its own private benchmark without losing ground. The answers are settling into shared standards: Databricks Omnigent gives one interface across Claude Code, Codex, Cursor, and Pi, OpenRouter's Fusion polls a panel of models and beats the frontier at half the cost, and enterprises now route hard tasks to strong models and easy ones to cheaper rivals. Portability is turning into the durable asset.
Answer trust (window closing)
Newer models score higher on tests but still make up answers, and the check layer is filling in fast. A Glean survey finds 69% of workers admit they ship agent output they never verified. A German court ruled Google liable for false answers in its AI Overviews. The fixes are arriving quickly: Axiom Math pushes proof-checked generation and hits 99% where an older model hit 5%, MIT and OpenAI train models to describe their own limits honestly, and LangSmith ships a cheap judge that flags where an agent went wrong. The window is closing as verification becomes standard.
Why subscribe
Windows do not stay open. The gap no one is solving today is the best opening on the board. The first real answer starts to close it. This month two gaps have no answer at all: no one can stop a model being copied once it is served, and no one has a trusted brake on AI that improves itself. Subscribe below to get next month's signals in your inbox, so you see the next opening before it closes.