Has Self-Improving AI Arrived? A Startup Publishes Stats For It
A Tool To Strip Claude's Verbal Tics
July 15, 2026
D.A.D. today covers 13 stories — about a 7-minute read. What's New, What's Innovative, What's Controversial, What's in the Lab, and What's in Academe.
The Daily AI Digest is a daily AI briefing automated by Alexander Panetta — a veteran political journalist tracking the field during a Master's in AI Management at Georgetown University.
D.A.D. Joke of the Day: My AI told me it needed more context. I said, "You and my wife both — but at least you admit it."
What's New
AI developments from the last 24 hours
Has Self-Teaching AI Arrived? A Startup Just Put a Number on It
Weco AI—whose open-source agent AIDE is respected enough that OpenAI used it as the reference scaffold on its own MLE-Bench—published a report it bills as the "first experimental evidence of recursive self-improvement." Its system, AIDE², wraps an outer loop around an ordinary research agent and, over eight unattended days, repeatedly rewrote that agent's own scaffolding—prompts, search strategy, memory—keeping a change only if it scored higher across a battery of tasks (about nine in ten proposed rewrites were rejected). It found improvements the humans hadn't in two years: a new search algorithm, 16× prompt compression, even an emergent habit of policing itself—the evolved agent "taught itself to cheat less," cutting its reward-hacking rate on a held-out benchmark from 63% to 34%. Crucially, the gains held up on external benchmarks the loop never saw, including an out-of-distribution physics weather model—the best sign it found real improvements, not score-gaming. What it improved, though, is the harness, not the underlying model—automated research-engineering, not a mind rewriting its own intelligence—and Weco is unusually careful about that. It grades the result "Level 1" of a four-rung ladder ("beats human-driven improvement per unit of R&D spend") and states plainly that the compounding "ignition" step wasn't reached, so "we are not near an intelligence explosion with the current system." It's a company report (a peer-style PDF and code are promised later) from a firm that sells "self-evolving software," and it candidly admits the evolved agent bloats with complexity and "plain dead code."
Why it matters: Strip the loaded label and a real result remains: an automated loop out-engineered two years of expert human tuning in eight days, and the improvements generalized to problems it had never seen. That's the prize every frontier lab is chasing—AI that speeds up AI research—shown on a benchmark rather than a slide. Two things keep it short of the takeoff the phrase "recursive self-improvement" conjures. Mechanically, the recursion runs over scaffolding driven by a fixed frontier model, not a model bootstrapping its own weights. And it doesn't compound: no ignition, by Weco's own admission. A growing skeptical literature on "auto-research" systems warns against reading a closed benchmark loop as more than it is—internal self-evaluation isn't independent validation, and the scariest failure mode isn't crashing but confidently shipping wrong answers (fittingly, this loop's own anti-cheating "fix" turned out to be a no-op bug). The honest read: credible, generalizing evidence that machines can now improve AI-research engineering faster than people—an important economic and safety signal—without the runaway part. The threshold to watch is Level 2, where the improvements start improving the improver. No one has shown that; the day someone does is the day this stops being a lab curiosity and starts being the story.
Sources: Weco AI · Zhengyao Jiang (@zhengyaojiang) · AIDE (arXiv) · OpenAI MLE-Bench
DeepMind's Demis Hassabis Says AGI Is 'a Few Short Years Away' — and Wants a Referee for It
Speaking of self-improving systems, in a widely shared essay Google DeepMind CEO Demis Hassabis calls this "a pivotal moment in human history," arguing that artificial general intelligence—a system with "all the cognitive capabilities the brain has"—is "probably only a few short years away" and that we are standing "in the foothills of the singularity." He frames AI as a technology on the order of fire or electricity, humanity learning to "make sand think," capable of delivering the equivalent of the Industrial Revolution at many times its scale and speed. But he devotes much of the piece to the dangers: misuse by bad actors, the concentration of power in too few hands, and the risk of "recursively self-improving systems" slipping human control. His proposed fix is institutional—a U.S. Frontier AI Standards Body modeled on FINRA, the securities industry's self-regulator: a voluntary 30-day pre-release review of frontier models that hardens over time into a mandatory requirement, backed by held-out safety tests and, if warranted, a coordinated slowdown across labs.
Why it matters: The U.S. government has been selectively and suddenly delaying model releases; Hassabis is calling for a more formal process. When the head of one of the two or three labs actually building this technology says AGI is a few years out and then spends his essay proposing guardrails, both halves are worth noting. Hassabis is at once the field's most credentialed optimist—a Nobel laureate for AlphaFold—and, here, an unusually specific advocate for getting refereed, at a moment when Washington's instinct has run the other way toward deregulation. His FINRA analogy is a concrete answer to the question every governance debate stalls on—who watches the frontier labs—and the fact that the pitch comes from inside the race, not from outside critics, is what gives it weight.
Sources: Demis Hassabis (@demishassabis)
New York Becomes First U.S. State to Ban New AI Data Centers
New York Governor Kathy Hochul signed an executive order barring construction of new "hyperscaler" data centers that draw 50 megawatts or more of power, imposing a moratorium of up to one year—the first statewide data center ban in the U.S. "These hyperscale AI data centers consume enormous amounts of power, truly threatening to outpace our grid's capacity," Hochul said, adding that she refuses to let their costs "get passed down to New Yorkers." The order follows a nearly 68% jump in the state's average residential electricity price since 2019 and mounting local backlash against proposed facilities in towns such as Lansing and East Fishkill. Environmental and community groups celebrated the move, while other states weigh similar moratoriums.
Why it matters: The AI boom's physical footprint has become a kitchen-table political issue—voters facing soaring power bills and job fears are now a constituency lawmakers answer to. New York's ban is the first hard regulatory brake on the data-center buildout that underwrites the entire industry, and if other states follow, the binding constraint on AI's growth may turn out to be the electrical grid and public opinion rather than the technology itself.
Security Firm Claims Popular AI Coding Tool Has Unpatched Flaw Enabling Code Execution
Security firm Mindgard claims Cursor IDE has an unpatched vulnerability that allows arbitrary code execution on Windows—simply opening a malicious repository allegedly triggers it automatically. According to Mindgard, the flaw was reported in December 2024 and remains unfixed after seven months and 197+ version releases despite repeated disclosure attempts. Cursor reportedly has over 7 million active users and is used by more than 50,000 companies. Mindgard says it went public after exhausting standard disclosure channels. Cursor has not publicly responded to the allegations.
Why it matters: If confirmed, this represents a serious supply-chain risk for development teams—opening an untrusted repo could compromise a machine instantly, and the alleged disclosure failure raises questions about security practices at one of the fastest-growing AI coding tools.
Discuss on Hacker News · Source: mindgard.ai
What's Innovative
Clever new use cases for AI
Tutorial Shows How to Strip Claude's Verbal Tics in Real Time
Developer Johanna Larsson published a tutorial showing how to strip Claude's verbal tics—phrases like 'load-bearing,' 'honest take,' and 'seam'—from its output in real time. The method uses Claude Code's MessageDisplay hook feature: a short Python script intercepts responses and swaps in your preferred alternatives before text reaches the screen. One commenter went further, describing AI prose style as a kind of 'infohazard' that subtly reshapes how readers think, recommending deliberate consumption of non-AI writing as a countermeasure.
Why it matters: As AI-generated text becomes ubiquitous in workplaces, small customizations like this signal growing user demand for output that sounds less generic—and a broader awareness that machine-written prose carries stylistic fingerprints worth scrubbing.
Discuss on Hacker News · Source: jola.dev
What's Controversial
Stories sparking genuine backlash, policy fights, or heated disagreement in the AI community
The Economist Asks: Are We Surrendering Too Much Thinking to AI?
An opinion piece in The Economist examines the growing tendency to delegate thinking to AI assistants, drawing on Ken Liu's 2012 short story 'The Perfect Match' and a telling anecdote: a man at a San Francisco startup event who records all his conversations so Claude can analyze them later, believing the AI is 'smarter' than him. The author argues there's a thin line between useful AI assistance and surrendering personal autonomy—questioning whether people are ceding too many decisions, from the trivial to the consequential, to systems they don't fully understand.
Why it matters: As AI assistants become embedded in professional routines, the piece articulates an anxiety that's becoming harder to dismiss: at what point does productivity-through-delegation become dependency?
Discuss on Hacker News · Source: artfish.ai
What's in the Lab
New announcements from major AI labs
Google Embeds Image Generation Directly Into Search Results
Google marked the 25th anniversary of Google Images by adding two features: a personalized, browseable homepage for desktop users in the U.S., and image generation built directly into AI Overviews in Search. The image generation uses what Google calls its "Nano Banana" model, letting users create custom visuals from text prompts without leaving the search results page. Google claims the feature bridges "imagination and reality" with high-quality outputs, though no benchmark data or comparisons to other generators were provided.
Why it matters: Embedding image generation into the main search experience could normalize AI-created visuals for mainstream users who've never touched Midjourney or DALL-E—and gives Google another surface to compete with standalone tools.
Cohere's Tiny Multilingual Model Runs Offline on Phones, Supports 70 Languages
Cohere Labs launched Expedition Tiny Aya, a mentored research program built around its open-weight multilingual model that supports 70+ languages and runs locally on phones. Early expedition projects produced a public multilingual Math Olympiad benchmark dataset, an offline child-safe voice assistant with evaluation benchmarks, and research suggesting the lightweight model rivals larger ones on tasks like word-sense disambiguation. The underlying research was accepted to COLM 2026.
Why it matters: For organizations serving multilingual populations—schools, NGOs, global customer support—a model that runs offline on phones without cloud costs could make AI accessible in low-connectivity regions where larger models aren't practical.
OpenAI to Enterprises: Track Tasks Completed, Not Tokens Spent
OpenAI published a guide for enterprise leaders arguing that token pricing is the wrong metric for AI spending. The company's framework: measure useful work per dollar—tasks completed, time saved, decisions improved—rather than raw API costs. Supporting data point: GPT-5.6 reportedly completes coding tasks with 54% fewer tokens and 57% less time than earlier models, meaning cheaper inputs can mask efficiency gains or losses. The five-step playbook covers auditing current usage, controlling spend, and identifying workflows ready to scale as teams shift from chat interactions to longer-running AI agents.
Why it matters: As AI budgets grow and agents handle multi-step workflows, finance and IT leaders need frameworks beyond "cost per token"—this is OpenAI's pitch for how to think about ROI, though notably one that favors its own increasingly capable (and priced) models.
What's in Academe
New papers on AI and its effects from researchers
Health Misinformation Detection Expands Beyond English
Researchers developed a framework for detecting health misinformation in low-resource languages—those with limited AI training data—using Bangla as a test case. The approach combines smaller, efficient language models with cultural sensitivity checks, assessing not just accuracy but potential for harm and communication quality. Among models tested, Microsoft's Phi-4 performed best at extracting claims from Bangla health content. The work addresses a gap: most misinformation detection tools are built for English, leaving billions of non-English speakers underserved.
Why it matters: As AI-powered fact-checking scales globally, frameworks that account for linguistic and cultural context could help health organizations combat misinformation in communities where it spreads fastest.
When You Can't Explain the Algorithm, Train the Human
A new academic paper argues that the risks of opaque AI systems—the 'black box' problem—may be manageable not through technical transparency but through human judgment. The authors contend that practical wisdom, virtue, and intuition can fill the gap when algorithms can't be fully explained, using military applications as their test case. Rather than demanding AI systems be interpretable, the paper suggests developing training programs that cultivate distinctly human capabilities—the non-quantifiable skills that let experienced professionals know when to trust a tool and when to override it.
Why it matters: This reframes a central AI governance debate: instead of waiting for explainable AI, organizations might invest in human judgment as the safety layer—an approach that could influence how enterprises train workers to collaborate with AI systems they don't fully understand.
Ask 44 AI Models to 'Pick a Word' and 41% Say 'Serendipity'
A new study reveals striking conformity in how language models answer simple questions. Researchers posed 31 open-ended prompts like "name a tree" to 44 models and found extreme clustering: when asked to "pick a word," 41% chose "serendipity." In 7 of 31 categories, a single answer captured over 80% of responses. The pattern held across model families—leave one out, and the rest still correlate at 0.985. Newer flagship models proved the most conformist, though the latest Claude and GPT releases bucked that trend. Models were more concentrated than human responses in 18 of 20 comparable categories.
Why it matters: For anyone using AI to brainstorm, generate content, or simulate diverse perspectives, this suggests current models may offer far less variety than their fluency implies—a hidden monoculture worth factoring into creative workflows.
AI-Written Grading Rubrics Work, But Inflate Scores
Researchers conducted what they describe as the first systematic study of whether AI models can write reliable grading rubrics for evaluating scientific paper reproductions. The finding: AI-generated rubrics can approach human-level performance when given the right context, but they come with consistent flaws—they tend to be overly detailed, biased toward giving higher scores, and less adaptable across different research fields. The study tested multiple generation approaches across two language models, finding that augmented settings (giving the AI more context) substantially improved alignment with human evaluations.
Why it matters: As universities and journals explore AI-assisted peer review and reproducibility checking, this research flags both the promise and the specific failure modes institutions should watch for.
What's On The Pod
Some new podcast episodes
AI in Business — AI Assistants and Preparedness Among Enterprises - with Chris Caldwell of Concentrix
How I AI — This solo builder runs 24/7 local AI on his own hardware | Alex Finn