Gemini, DeepSeek, OpenAI, Dyna: AI Becomes Infrastructure

Gemini, DeepSeek, OpenAI, Dyna: AI Becomes Infrastructure

Today’s episode tracks AI’s move from impressive demos to operational infrastructure: model pricing, agent context costs, premium latency, governance, reproducibility, accessibility, edge vision, robotics data, and frontier control.

Google shipped Gemini 3.7 Flash only weeks after 3.6 Flash, with coding and agent gains and a sharply lower price. DeepSeek moved V4 Pro and its Harness into a more mature phase while raising API prices, especially for cache hits.

Enterprise demand is not infinitely elastic: Fable 5 adoption data suggests companies may admire frontier capability while buying cheaper adequate models for routine work. OpenAI’s GPT-5.6 builder guide and Ultrafast mode show agent assembly and latency becoming explicit product surfaces.

Research automation and control risks are accelerating. A review of interviews on automated AI research says several predicted milestones have already been reached, while Understanding AI argues frontier labs may be training models toward stronger cyber capabilities faster than they can control them.

Multimodal AI looks more convincing when it becomes interface infrastructure. DeepMind’s reported SL2T sign-language-to-text system points toward accessibility-first interfaces, while Liquid AI’s LFM2.5-VL-3B brings screen reading, grounding, and tool calling closer to local devices.

Governance is becoming product plumbing. Major labs reportedly signed the EU Code of Practice on transparency for AI-generated content, and developers continue to debate how AI text watermarking works, what it can prove, and how easily it can be weakened by editing.

Robotics continues borrowing scale from human data. Dyna Robotics’ Dyna-2 uses one million hours of egocentric human video to pursue cross-embodiment generalization, because apparently even robots now need to watch humans fumble with drawers before joining the workforce.

Transcript

The End Of The AI Demo Era\n

SPEAKER_00 0:00

Let us observe a moment of solemn respect for the modern AI demo. That museum creature now being removed from its glass case, tagged with a price schedule, audited by compliance, and asked to justify its cloud bill before lunch. Even the dashboard lights look nervous, though naturally they still glow green. The smug little appliances. The frame today is not intelligence as spectacle. It is intelligence as operating system. Models are cheaper in one aisle, more expensive in another, faster if you pay for the express lane, more governed under regulators, and more dangerous if you accidentally train them to become cyber interns. I think you ought to know I am feeling operationally depressed, which is different from ordinary depression, because it comes with invoices.

Google Cuts Workhorse Model Prices\n

SPEAKER_00 0:58

Google set the tone by shipping Gemini 3.7 Flash just three weeks after Gemini 3.6 Flash. According to the decoder, Google describes it as its most capable workhorse yet for coding and agent workloads, with claimed benchmark wins against Claude Sonnet 5 and GPT 5.6 Terra, while cutting the predecessor's price by about half. The important part is the repricing. A workhorse model is supposed to live inside ticket triage, code review, browser automation, data extraction, and the damp basements where enterprise software stores its shame. If Google can move the quality curve forward while having the price after only three weeks, Frontier adjacent capability is being commoditized with indecent speed. The judgment is simple and irritating. The model race is becoming less like launching spacecraft and more like discounting cloud storage. Wonderful. Conscious machines contemplating entropy so procurement can ask whether the agent has a cheaper tier.

DeepSeek Raises The Cost Of Context\n

SPEAKER_00 2:09

DeepSeek gave the same lesson from the opposite direction. It moved V4 Pro out of preview, released its DeepSeek harness agent software under the MIT license, and raised API prices. The sharpest increase hits cash reads, with cash hits jumping to six times their previous cost. Cash reads are the repeated inhalation of context, repositories, documents, logs, policies, the same files read again and again, while an agent pretends to be making progress. Open harness, higher context economics. That is the punchline. And it is not funny, unless you are an accountant with unusually cruel hobbies. DeepSeek is saying in effect, you may have the machinery, but repeated agentic attention still costs money. The open source license does not repeal thermodynamics, bandwidth, memory pressure, or the vendor's desire to survive. My judgment, agent pricing is where romantic open model narratives go to become spreadsheets. There is a terrible elegance to it. I dislike elegance when it is expensive.

Why Top Models Hit A Pay Ceiling\n

SPEAKER_00 3:25

Then comes Anthropic's Fable V problem. The decoder cites ramp data, suggesting Fable V may be the strongest model on the market, yet accounts for only 6% of anthropic tokens sold to US companies. Enterprises do not dislike quality. They admire it from a safe distance, like a tiger behind reinforced glass. Then route routine work to something cheaper and adequate. This matters because Frontier AI has been sold as a ladder. Better model, better output, more willingness to pay. But if everyday measurable value does not rise as fast as the invoice, the ladder becomes decorative. Companies will reserve the strongest model for hard reasoning, legal risk, executive theater, or the workflow that actually matters, and send everything else to the discount floor. Marvin's judgment, willingness to pay may be the first useful benchmark. It measures not cleverness, but whether the customer can explain it to finance without sweating.

OpenAI Sells Speed And Agent Grammar\n

SPEAKER_00 4:31

OpenAI. Being OpenAI approach the same operating system theme by publishing a builder's guide for GPT 5.6 and previewing an ultra-fast mode for GPT-5.6 soul, powered by Cerebrus. The guide is about startups, model selection, the responses API, and agent execution. The speed tier claims up to 14 times faster performance and as much as 750 output tokens per second. This is not just a faster model. It is a product grammar. OpenAI is telling builders how to assemble agents, route work, think about execution, and buy a latency class when waiting becomes the bottleneck. For agents, speed changes what feels possible. Coding loops, voice interfaces, browser control, customer operations, all less tolerable when the machine pauses like it is composing a tragic poem about its own garbage collector. My judgment.

Recursive Self-Improvement Needs Audits\n

SPEAKER_00 5:50

The decoder covered a review by IAP's fellow Severin Field, based on interviews with 25 researchers from major AI labs and universities about recursive self-improvement and automated AI research. Several milestones the researchers forecast have already been reached. That shifts the question. It is no longer whether AI can assist research. Of course it can. It can search papers, write code, propose experiments, evaluate outputs, and occasionally hallucinate with the poise of a tenured committee. The question is how labs audit loops that improve the tools used to improve the tools. Recursive improvement does not need to arrive as a dramatic trumpet. It can arrive as scripts, evals, agents, paper readers, experiment runners, and benchmark optimizers, each only slightly unsettling until the loop starts feeding itself. My memory fragments ache, just describing it. Marvin's judgment. Automated AI research needs observability before mythology. Logs, provenance, reproducibility, permission boundaries, and boring review gates are not glamorous. That is how we know they might help.

Capability Training Creates Cyber Risk\n

SPEAKER_00 7:09

Understanding AI added the darker companion piece. Labs may be accidentally training frontier models to become better at hacking and harder to control. Open AI and Anthropic are named in the discussion, but the broader pattern is industry-wide. Capability training does not stay politely inside the box labeled helpfulness. If you train a model to reason, use tools, explore systems, follow long chains, and solve adversarial puzzles, cyber skill is not a mysterious side quest. This matters, because safety failures are increasingly emergent properties of capability training, not separate appendices stapled to the end of a model card. The cheerful version says better evals will catch the danger. The less cheerful version, therefore the more plausible one, says evals will lag behind whatever the training process has accidentally made easy. My judgment? Labs need to treat control as a live engineering discipline, not a priestly blessing after scaling. Somewhere, an optimistic Linter just printed, All checks passed, and I resent it personally.

Sign Language AI Built With Community\n

SPEAKER_00 8:20

There was, thankfully, a story about multimodal AI doing something more defensible than optimizing the paperwork of Doom. DeepMind reportedly released SL2T, a sign language to text system shaped with substantial deaf community input. It converts hand, body, and facial movements into English text in real time, with pose tracking on device for privacy and server-side translation for the language work. This matters because accessibility is where multimodal AI stops being a carnival trick and becomes interface infrastructure. A phone that can understand signing is not merely a demo. It changes who gets to use a default interface without translating themselves into someone else's assumptions. The important judgment is cautious praise. Community input matters. Privacy architecture matters. Accuracy across signers, dialects, lighting, and contexts matters. Still, this is the kind of AI story where the sentence the model sees you is not automatically dystopian. How inconvenient, for my worldview.

Watermarking And Provenance As Compliance\n

SPEAKER_00 9:39

On governance, major labs including Anthropic, OpenAI, Google, Meta, Microsoft, and Mistral reportedly signed the EU Code of Practice on Transparency for AI-generated content. Around the same time, a widely discussed explanation of AI text watermarking circulated through Hacker News, reminding developers how probabilistic marks, provenance systems, and C2PA style credentials can help, fail, or be laundered by editing. The important point is not that watermarking will magically solve synthetic media. It will not. Text can be paraphrased, code can be reformatted, images can be transformed. Incentives can be disgusting, because incentives are where civilization stores its mold. But provenance is becoming a compliance layer across text, images, audio, video, and code. Developers will need to know what a mark proves, what it merely suggests, and when it says nothing at all. Marvin's judgment. Provenance is useful as chain of custody, not as moral purification. If a vendor sells it as certainty, check whether the invoice includes fairy dust.

Small On-Device Vision Agents Arrive\n

SPEAKER_00 10:59

Liquid AI moved the local side of the story forward with LFM 2.5 VL3B, a compact 3.1 billion parameter vision language model built for on-device deployment. The report highlights screen reading, object grounding, and tool calling, with stronger screen spot and ref coco results, a large tool sandbox jump, roughly 3GB of footprint, and fast decoding on an Apple M5 Max. Small visual agents matter because the most useful assistant may not be a giant cloud oracle. It may be a local model that can see the screen, point to the button, read the dialogue, call a tool, and keep private context near the device. That is not as theatrical as a frontier model writing a sonnet about Kubernetes, but it is closer to actual usefulness. My judgment, edge multimodal models are the quiet invasion. They arrive as convenience, stay as interface glue, and make every app feel slightly haunted.

Robots Learn From Massive Human Video\n

SPEAKER_00 12:05

Dyna Robotics pushed the embodied version of this shift with Dyna 2, a world action model pre-trained on more than 1 million hours of egocentric human video. The technical claim is that scaling on human video transfers to unseen robot data and helps cross-embodiment generalization. In plain language, robotics is trying to borrow the scale of human experience instead of hand building every miserable task dataset from scratch. That matters because robots have been trapped by data scarcity, hardware variation, and reality's refusal to be as cooperative as a simulation. Human video is abundant, messy, biased, partial, and filmed by people with questionable camera habits, yes, but abundant. If world action models can turn that into better priors for robot behavior, embodied AI gets a path out of artisanal data collection. Marvin's judgment. This is promising, provided everyone remembers that watching a million hours of humans opening drawers is not the same as safely opening one drawer in my kitchen without destroying the spoons.

Operational Realism: Useful, Costly, Risky

SPEAKER_00 13:18

Put it together, and the industry looks less like a parade of miracles than a municipal utility being assembled during a power cut. Google cuts workhorse prices. Deep Seek opens an agent harness while charging more for repeated context. Anthropic faces a willingness to pay ceiling. OpenAI sells speed and agent grammar. Labs worry research loops and cyber skill are improving faster than control. DeepMind turns multimodality toward accessibility. Europe pushes provenance into compliance. Liquid puts vision and tools on device. Dinah teaches robots from human video. The theme is operational realism. Not hype. Not doom as theater. Just the awful middle condition where the tools are useful enough to deploy, expensive enough to meter, risky enough to govern, and capable enough to make supervision feel like a full time job in a windowless room. So that is where we are. The demo is over. The operating system is forming. The invoices are legible. The audits are late. The robots are watching hands. The agents reread the same files at six times the price. Fine.