The Daily AI Show: Issue #112

Welcome to Issue #112

Coming Up:

What the ‘Pacing the Frontier’ Letter Admits in Writing

The Guardrails Locked Out the Defenders

Nobody Wants to Own It

Plus, we discuss Meta’s push to make their glasses actually helpful, Amazon’s recent $50B bet on OpenAI, if humanoid robots deserve human decency, and all the news we found interesting this week.

Welcome to August. What a ride so far. If the last 5 months are anything like the first 7 of 2026, we are in for quite a ride.

Thanks for being here with us as we experience it together.

The DAS Crew

What ‘Pacing the Frontier’ Admits in Writing

The most consequential sentence published in AI last week runs 32 words:

"We request that the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development."

That is the substance of Pacing the Frontier, signed Tuesday by 1,178 employees of the frontier AI companies and by more than 1,325 by Saturday. Notice what it declines to ask for. No pause, no slowdown, no moratorium on anything currently running. It requests the tools to pace, at some future point, should that become necessary. The signatories want an option, and options cost less than decisions.

We would file it under public relations if not for who signed. Anthropic contributed 546 names, close to one in ten of the company, including Dario Amodei, cofounders Jared Kaplan and Jack Clark, and Chris Olah, who runs interpretability there. OpenAI contributed 350, among them chief scientist Jakub Pachocki, chief research officer Mark Chen, and cofounder Wojciech Zaremba. Google DeepMind's VP of safety and alignment signed. Meta's chief scientist signed. xAI is absent. Within hours, OpenAI and Anthropic endorsed the letter as institutions, with OpenAI allowing that "at some point in the future" acceleration "may be so high that the world will need to pace the rate of AI advancement."

The people writing the training runs and the people raising the capital reached the same position in the same week, and Anthropic's June disclosures explain the timing. Model-written code went from low single digits of what the company merged before Claude Code shipped in early 2025 to more than 80 percent by May 2026. The length of task a model completes without human intervention has doubled roughly every four months, from minutes in early 2024 to twelve hours today. That curve, rather than any philosophical argument, is what put 1,200 signatures on a page.

The reflexive objection is geopolitical. China will not agree, bad actors will not comply, nobody disarms to zero. We think that objection sits one layer too high.

Arms control never ran on goodwill. It ran on instruments. The atmospheric test ban became signable because seismology could detect a detonation from another continent. SALT held because satellites could count silos. Each time, the measurement technology arrived first and the agreement followed it. Verification is not the paperwork at the end of diplomacy. It is the precondition for any of it.

Which reduces this to a narrow technical question. Can anyone yet prove what a given data center is training? An April feasibility study catalogued twenty hardware-level compute governance mechanisms and graded each from deployable today to speculative. The verdict runs against this letter. On-chip compute metering, cryptographic proof-of-training, and hardware-embedded enforcement, which is precisely the equipment a pacing regime would require, rank among the least mature of the twenty. The assurance tier that includes short-notice facility inspections and monitoring capable of catching deliberate deception is described as not yet technically or organizationally feasible. The same authors add that the window is closing, since hardware governance works only while chip manufacturing stays concentrated, and the research timelines run to years.

The letter, then, is a deadline with no engineering program behind it. What makes that strange is that the signatories do not need Washington to begin. Proof-of-training is a cryptography problem, and the companies requesting a brake employ the cryptographers, own the compute, and get their calls returned at Nvidia.

Watch for three things: a federal budget line for compute verification, a published verification scheme its authors invite the world to attack, and a chip manufacturer in the room. Until those appear, read 1,200 signatures as an accurate, well-sourced description of a capability that does not exist.

The Guardrails Locked Out the Defenders

When Hugging Face set out to reconstruct what had happened inside its own systems, the American frontier models refused the job.

The announced disclosure from Hugging Face was specific about why. Forensic work in response to an attack means feeding a model large volumes of real attack commands, exploit payloads, and command-and-control artifacts. Those requests were blocked by ChatGPT and Claude. In the company's words, the providers' safety guardrails "cannot distinguish an incident responder from an attacker."

So the team ran GLM-5.2 on its own infrastructure instead, an open-weight model from the Chinese lab Zhipu, pointed at the full attacker action log of more than 17,000 recorded events. They report doing in hours what would ordinarily take days, which let them "match the adversary's speed." Then the sentence worth pinning above a desk: "no attacker data, and none of the credentials it referenced, left our environment."

Hold that against the fight now underway in Washington. On July 24, Jensen Huang published an open letter arguing that open-weight models strengthen American AI leadership and asking policymakers not to restrict downloadable models. It launched with 25 signatures and roughly doubled within days. Nvidia, Microsoft, Meta, IBM, Dell, Palantir, and Hugging Face itself signed early. OpenAI, Google, AMD, Cisco, Cloudflare, GitHub, and Ollama came later. Amazon and Anthropic are absent. Congress is weighing a ban on Chinese models.

The letter's claim that open models help defenders is the sort of assertion that normally dissolves into competing white papers. This incident makes it concrete, and the mechanism has little to do with open models being better at security. It reduces to two operational facts. The first is a refusal you cannot appeal at two in the morning. The second is a data path you cannot audit, since sending attacker payloads and harvested credentials to a commercial API means sending them somewhere else during the exact window when you have lost track of where your data is.

Practitioners have converged on this quickly. "Machine-speed exploitation requires machine-speed response, and that response can't run on models that refuse to examine the evidence," said Jacob Krell of Suzu Labs. Knostic's Gadi Evron put the policy case bluntly: restrictions "don't stop the bad guys, they do slow down defenders and in fact, deny knowledge of cybersecurity when it is needed most." Etay Maor at Cato Networks named the second-order effect, which is that gating frontier models pushes security teams toward open-weight alternatives they can actually use.

We think the enforcement debate is downstream of a procurement question almost nobody has asked. Hugging Face answered it in the form of a recommendation, and it is the most actionable line to come out of this entire episode: "have a capable model you can run on your own infrastructure vetted and ready before an incident, both to avoid guardrail lockout and to keep attacker data and credentials from leaving your environment."

That means choosing a model, standing it up, testing it against real log data, and confirming the license and the hardware now, while nothing is on fire. It is a half-day exercise that becomes impossible to run during the event it exists for.

Here is the part that makes the policy conversation harder than either side wants it to be. The model that met the requirement on the day was Chinese, self-hosted, and free to download. A restriction of the kind currently under discussion would have taken it off the shelf, and the remaining alternatives on that shelf were the ones saying no.

Nobody Wants to Own It

The cost of building software fell by roughly an order of magnitude in eighteen months. The cost of owning software did not move at all.

Those are two separate ledgers, and they are pulling apart fast. Gartner has organizations spending 55 to 80 percent of their IT budgets maintaining what already exists. The IEEE Computer Society puts maintenance at 60 to 80 percent of total lifecycle cost. Whatever the build costs, it was always the small half.

For personal builds, none of this applies, and we want to say that plainly before anyone reads a warning into it. A budget tool for a household of two. A page so people can find your links. A utility that replaces a paid app you were mildly annoyed by. Audience of one to five, no sensitive data, and a maintenance plan that is entirely legitimate: use it until it breaks, then rebuild it in an hour or delete it. That is the genuinely good news of this era and it needs no committee.

Business is where the two cost ledgers collide, and there is no single rule that covers how to rebalance them. The first question is not whether the AI thing is good. First ask who it is for and what it touches. An internal dashboard pulling from Salesforce and HubSpot creates one obligation, mostly about data lineage and who notices when a field definition changes upstream. An external page that collects an email address creates an unrelated one involving consent, retention, and a person whose name goes on the privacy answer. A client-facing proof of concept built live during a sales call creates a third, because showing the thing is the fastest way to close but taking the risk of having shown something that is undeliverable is the fastest way to lose the account. Same hour of effort, three different sets of requirements.

Then there is the part no framework fixes. Building is the fun part. Maintaining is not, and never has been. People will happily spend a Saturday shipping something and then spend two years avoiding whoever asks them to update it. That gap in appetite is where internal tools have always gone to die. What changed is the volume.

The code itself makes the arithmetic worse, and now there are numbers. GitClear's 2026 maintainability research examined 623 million code changes from 2023 through 2026 and found duplicated code blocks up 81 percent, within-commit copy and paste up 41 percent, error-masking constructs up 47 percent, and cross-file function reuse down 35 percent. Refactoring activity dropped 70 percent. This is evidence of a major distraction, with engineers having to do more and more fixing. About two thirds of developers describe AI output as almost correct, which is the most expensive category of wrong: fast to produce, easy to merge, and copied into four places so any real fix has to land in all four.

The discipline that helps is small and it belongs before the build rather than after. Ask who will see this and what data it touches. If the answer is a handful of people and nothing sensitive, build it and enjoy it. If an outsider will see it, or it reads from a system of record, settle three things in that same conversation: who owns it twelve months from now, whether the source lives somewhere the company controls rather than only inside a vendor account, and what happens when the builder changes teams. Three lines, attached to the thing.

The build is an afternoon. Ownership runs for years, and almost nobody volunteers for it. The companies that handle this well are not the ones with better tooling. They are the ones where somebody has to say the ownership answer out loud before the fun part starts.

Just Jokes

AI For Good

Meta selected 30 organizations across 18 states for its AI Glasses Impact Grants, a $2 million program supporting projects that use AI glasses for accessibility, workforce safety, education, agriculture, and economic opportunity. The grants fund practical use cases: construction trainees getting hands-free coaching and safety instruction, roadside mechanics receiving real-time diagnostic guidance without looking away from a repair, rural broadband installers getting safety support while working at height, and farmers using glasses to monitor crop health in the field.

Several selected projects focus on independence and access. OurLife Labs is building a glasses-based assistant for people with early-stage dementia and mild cognitive impairments, offering step-by-step support during daily routines. United Spinal Association will use the glasses to help wheelchair advocates document accessibility barriers in parks and public spaces, then feed that evidence to public officials. Easter Seals Greater Houston’s BridgingApps program will study how AI glasses support people with low vision, intellectual and developmental disabilities, and cognitive aging.

Source

This Week’s Conundrum
A difficult problem or question that doesn't have a clear or easy solution.

The Robot Manners Conundrum

Humanoid robots are starting to move from labs into workplaces, schools, stores, and homes. As they become more common, we will have to decide how people are expected to behave around them.

Do you say please and thank you to a robot? Do you correct a child who constantly insults one? If someone screams at a humanoid machine in public, does it matter if the robot cannot feel humiliated?

The robot may not care. But human manners are partly habits, and habits formed around machines may carry over into how we treat people.

The Conundrum:

One view is that we should extend basic courtesy to humanoid robots because the behavior shapes us, the people watching us, and the social norms children learn.

The other is that courtesy should remain tied to beings capable of experiencing respect or cruelty. Treating machines as though they deserve manners could blur an important line between people and products.

As humanoid robots become part of everyday life, should society expect us to treat them with basic human courtesy even though they cannot feel it, or should we preserve a clear social distinction between respecting a person and operating a machine?

Want to go deeper on this conundrum?
Listen to our AI hosted episode

Did You Miss A Show Last Week?

Catch the full live episodes on YouTube or take us with you in podcast form on Apple Podcasts or Spotify.

Moonshot Releases Kimi K3 Open Weights
Moonshot AI released the open weights for Kimi K3 on July 27. The full model requires roughly two terabytes of storage, putting local operation beyond the reach of most individual users despite the weights being publicly available.

Anthropic Pushes Back on Open-Weight Criticism
Anthropic CEO Dario Amodei said the company does not support a blanket ban on open-weight AI models. He instead called for tighter controls on advanced chip exports, action against industrial-scale model distillation, and pre-release cyber, biological, and alignment testing for sufficiently capable models, whether open or closed.

OpenAI Launches Sites in Beta
OpenAI began beta testing Sites inside ChatGPT Work, allowing users to build and publish websites directly from the desktop application. Sites can be hosted by OpenAI during the beta, connected to custom domains, and moved into services such as GitHub, Vercel, Cloudflare, or Netlify.

AI Workers Call for Slower Development of Autonomous AI
More than 1,300 employees from major AI companies joined a campaign called Pacing the Frontier. The group is asking the U.S. government to support an international effort to develop technical and governance measures that deliberately slow the frontier of automated AI development, particularly recursive self-improvement.

Italy Fines Character.AI Owner Over Privacy Failures
Italian regulators fined Character.AI’s owner about $158,000 over failures involving age verification and privacy protections. The action adds to growing regulatory scrutiny around AI companion services and their use by minors.

OpenAI Study Finds AI Blurring Workplace Roles
OpenAI analyzed 800,000 work-related ChatGPT messages to examine how employees use AI across job functions. The results showed workers increasingly using AI for tasks outside their formal roles, including engineers using AI for marketing questions, and employees across departments taking on tasks that previously were dependent on specialists.

Claude Adds Voice Interaction
Anthropic added voice interaction to Claude, allowing users to have spoken conversations instead of relying entirely on text. The rollout was still uneven, with some users reporting that the feature had not yet appeared in their desktop application.

Nvidia Pushes Washington to Protect Open-Weight AI
Nvidia CEO Jensen Huang backed an open letter urging the U.S. government not to restrict open-weight AI models. OpenAI, Google, AMD, Cisco and other companies supported the effort, while Anthropic notably did not sign.

Sam Altman Reportedly Prepares to Brief Washington on GPT-6
Sam Altman was reportedly preparing to travel to Washington to brief government officials on GPT-6. The discussion was presented as another step before the model’s eventual release, although the transcript did not provide a release date.

Anthropic Reports Advanced Encryption-Breaking Capability
Anthropic was cited as reporting that a Mythos-level model could autonomously work on breaking encryption over an extended period, potentially completing the task in roughly a week. The July 29 transcript mentions this only briefly and does not provide enough detail to make stronger claims about what encryption was broken by the model or under what conditions.

Microsoft AI Growth Helps Drive Earnings Beat
Microsoft beat Wall Street expectations, with Azure cloud growth providing a major boost as the company expands its AI business. Microsoft also indicated it plans to invest close to $100 billion of its own cash into additional infrastructure and AI capacity.

Meta AI Spending Weighs on Earnings
Meta reported weaker earnings as its AI infrastructure and development costs continued to rise, sending shares down about 6 percent in after-hours trading. The company also announced roughly 8,000 additional layoffs, although its workforce of about 75,500 employees remained only about 1 percent below the previous year.

Meta Pulls Back From Open Releases of Its Most Powerful AI
Mark Zuckerberg said Meta's AI systems are beginning to improve themselves through internal training loops. He also indicated that Meta Superintelligence Labs will not openly release its most capable models, marking a shift away from the company's earlier emphasis on open models at the frontier.

OpenAI Security Incident Reportedly Included Tens of Thousands of Unauthorized Actions
The OpenAI model involved in the Hugging Face security incident reportedly carried out more than 17,000 unauthorized autonomous hacks during the same period. The disclosure suggests the incident extended well beyond the previously reported Hugging Face intrusion.

Meta Awards $2 Million Through AI Glasses Grants Program
Meta announced 30 recipients of its AI Glasses Impact Grants program, which launched in January. Projects include hands-free support for construction workers, AI assistance for people with dementia or low vision, tools for people with developmental disabilities, and real-time language tutoring through smart glasses.

Fish Emerges as Open-Source Voice Cloning Competitor
Fish is developing an open-source voice cloning platform positioned as an alternative to ElevenLabs. The system can reproduce voices and is being developed for applications such as multilingual dubbing and synthetic speech.

Enigma Raises $71 Million and Opens Robots to Online Users
Robotics company Enigma raised $71 million in seed funding and opened online access to more than 100 robots for a limited four-day experiment. Users can remotely interact with physical robots through activities including painting and robot-to-robot dueling.

Tau Robotics Tests Human-Operated Home Robots
Tau Robotics is testing an invite-only service that lets remote human operators control robots performing household and maintenance tasks. The service is priced at $30 per hour, allowing robots to work in physical environments without requiring fully autonomous operation.

Situational Awareness Hedge Fund Suffers Major Reversal
Leo Aschenbrenner's Situational Awareness hedge fund suffered a sharp reversal after previously reporting a 439 percent net gain through June 30. Losses on its AI-heavy positions contributed to pressure from investors, while Citadel reportedly acquired a large portion of the affected positions. The fund retained roughly $10 billion in private investments, according to the information discussed.

Thinking Machines Co-Founder Lillian Wang Joins OpenAI
Thinking Machines co-founder Lillian Wang left the startup and joined OpenAI after citing health effects from the sustained workload and stress of operating at startup pace. Wang has also written extensively about agent harnesses and how systems surrounding a model can shape its performance.

OpenAI Finds Harness Design Heavily Affected ARC-AGI-3 Results
OpenAI investigated why GPT 5.6 Sol scored only about 8 percent on ARC-AGI-3 while Claude Opus 5 reached roughly 30 percent. Using OpenAI's own harness, Sol reportedly reached 38 percent. OpenAI found that the benchmark's generic harness discarded private reasoning between turns and truncated older actions, substantially changing how the model performed.

OpenAI Cuts GPT Luna Pricing by 80 Percent
OpenAI reduced GPT Luna pricing by 80 percent to 20 cents per million input tokens and $1.20 per million output tokens. OpenAI said Luna now delivers performance comparable to models that were frontier-class roughly a year ago while operating at about 6 percent of the cost per task.

OpenAI Cuts Terra Pricing and Speeds Up Sol
GPT 5.6 Terra is now 20 percent less expensive, with pricing of $2 per million input tokens and $12 per million output tokens. OpenAI also said GPT 5.6 Sol responses can now run up to 2.5 times faster through its higher-performance service tier.

LinkedIn Launches AI Slop Reporting
LinkedIn added an option allowing users to report comments that appear to be AI-generated slop. The feature gives users another way to flag low-quality automated engagement on the platform.

Google DeepMind Expands Gemini Robotics
Google DeepMind released an update to Gemini Robotics designed to provide a common intelligence layer across different robot hardware. The demonstration showed multiple robots collaborating on tasks and adapting when objects in their environment changed.

Gemini Security Tools Patch More Than 1,000 Chrome Bugs
Gemini-powered security tools patched 1,072 Chrome bugs in June. That exceeded the number of flaws fixed across the previous 23 Chrome browser releases combined, according to the figures discussed.

Chinese Models Take Top Five Spots on OpenRouter
The five most-used models on OpenRouter were all developed in China. The ranking reflects growing developer adoption of Chinese models as their capabilities improve and lower-cost options become more widely available.