IMPLICATOR
San Francisco | Tuesday, May 12, 2026
Thinking Machines Lab is trying to move AI past the chat box by making timing part of the model. Its 200-millisecond response chunks sound small, but they point to the next interface fight: models that can interrupt, listen, watch video and stay useful while work is already moving.
Nadella's trial testimony pulls the same argument into corporate control. Musk says Microsoft captured OpenAI. Nadella says Musk had his number and never called.
Google adds the harder edge: AI may already be helping criminals find zero-days. Today's stack is not just smarter. It is faster, closer to the user and harder to contain.
Stay curious,
Marcus Schuler
Know someone drowning in AI noise? Forward this briefing. They can subscribe free here.
Thinking Machines Puts 200ms at Center of AI
Thinking Machines Lab is making the pause, the interruption and the half-finished thought part of the model. The old chat loop now looks slow by design.
Mira Murati's lab previewed interaction models that process input and output in 200-millisecond chunks across audio, video and text. The point is not a nicer voice wrapper. It is a model that can listen while the user is still working, correct course midstream and handle timing as learned behavior.
The early claim comes with caveats. TML-Interaction-Small is company-reported, and the real product test is ordinary work: hesitation, correction, multitasking and video context. But the direction is clear. The frontier fight is moving from smarter answers to better turn-taking.
Why This Matters:
- AI products that can listen and interrupt well will feel useful before users inspect the benchmark chart.
- Model labs are moving interaction quality from interface plumbing into architecture, where rivals cannot copy it with a UI patch.
Reality Check
What's confirmed: Thinking Machines published the research preview and says its model works in 200-millisecond chunks across audio, video and text.
What's implied (not proven): Better timing may become a durable product edge, not merely a demo effect.
What could go wrong: Low latency can still fail if the model interrupts badly, misses context or feels too eager.
What to watch next: Whether Thinking Machines turns the preview into a product people use outside controlled demos.
The One Number
20x - The reported demand for Cerebras' IPO versus shares available, according to Reuters. That is the cleanest market signal yet that investors want direct AI-chip exposure beyond Nvidia. Scarcity, not proven public-company operating history, is setting the tone before CBRS reaches the tape.
Source: Reuters, May 10, 2026
Nadella Says Musk Never Called on OpenAI Bet
Satya Nadella did not make Microsoft's OpenAI problem disappear. He made Musk's theory depend on the call that never came.
On the stand in Oakland, Nadella said Musk never contacted him about concerns that Microsoft's OpenAI investment violated special terms. The line that stuck was blunt: "We have each other's phone numbers."
That matters because Musk's case needs Microsoft to look less like a commercial partner and more like a knowing participant in OpenAI's alleged turn away from its nonprofit mission. Nadella's testimony left the money problem intact. It made the warning record look thinner.
Why This Matters:
- Microsoft's $13 billion OpenAI relationship remains central to the case, but testimony now favors the missing-objection defense.
- The trial keeps turning AI governance into a record-keeping dispute, where old calls, emails and board notes carry the argument.
AI Image of the Day
Prompt: elegant young woman in side profile, tiny blue birds resting gently on her fingers and shoulders, birds separate from the body, extremely delicate blue ink line art, ivory paper background, soft pastel blue watercolor accents, airy negative space, ultra fine sketch details, poetic atmosphere, graceful hand drawn fashion illustration, minimal color palette --ar 4:5 --sref 1352915912 --v 8.1
Google Finds AI-Assisted Zero-Day Exploit
Google says criminals likely used AI to help build a zero-day exploit. The attack was stopped, but the warning moved from theory to case file.
The reported exploit targeted a two-factor authentication bypass in an unnamed open-source administration tool. Google said the actor planned a mass campaign, but a vendor patch landed before the attempt could scale.
That is the new security story. AI does not need to make every attacker brilliant. It only needs to shorten the path from bug discovery to working exploit. Defenders now have the same tool class, which turns the patch window into the battlefield.
Why This Matters:
- AI-assisted bug hunting is now a practical risk for defenders, not only a frontier-lab safety debate.
- The strongest defense may be faster disclosure, triage and patching before mass exploitation begins.
🧰 AI Toolbox
How to Turn a Product Sketch or Reference Image into a Usable 3D Model with Tripo AI
Tripo AI turns text prompts or single reference images into textured 3D assets you can preview, refine, rig and export for Blender, Unity, Unreal or 3D printing. It is useful when you need a fast product mockup, game prop, printable prototype or rough environment asset but do not want to model from scratch. Browser-based with free credits and paid export options.
Tutorial:
- Go to Tripo AI and sign in.
- Choose "Image to 3D" for a sketch, product photo or AI-generated reference, or "Text to 3D" for a clean prompt.
- Upload one clear reference image with the object centered and minimal background clutter.
- Generate the first model, then rotate the preview and check the silhouette, texture and obvious geometry errors.
- Use refine, texture or style controls to improve the asset before exporting.
- If it is a character or creature, run auto-rigging so the model can be posed or animated later.
- Export as GLB, OBJ, FBX or STL, then clean scale and topology in Blender before using it in production or printing.
What To Watch Next
|
||
|
||
|
💡 5-Minute Skill
Tuesday, 5:42 p.m. The launch channel has 184 messages, three emoji polls, one half-decision and a thread where someone says "works for me" without saying what works. You are not behind because you missed the meeting. You are behind because the meeting kept happening in Slack after the meeting ended.
Your raw input:
Channel dump from #launch-check, May 6-12. Beta email moved from Wednesday to Thursday, maybe final. Priya says pricing page copy is blocked on legal approval. Sam says the demo video is "basically done" but needs product screenshots from Leo. Two customers asked whether SSO is included in beta. Jenna reacted thumbs-up to "existing trial users only" but nobody wrote it as a decision. Open question: whether EU customers are excluded until DPA language is approved. Deadline: newsletter mention Friday morning.
The prompt:
Act like a chief of staff cleaning up a launch channel. Do not summarize chronologically. Extract: confirmed decisions, likely decisions that need confirmation, blockers, owners, deadline risks and open questions. Flag anything with no named owner. End with one Slack message I can post to confirm the decisions and force owners without sounding like a scold.
The output:
Confirmed decisions: none explicitly written. Likely decisions to confirm: beta email moves to Thursday; beta audience is existing trial users only. Blockers: pricing page waits on Legal; demo video waits on Leo's screenshots; EU beta status waits on DPA language. Owners missing: Legal approval owner, EU decision owner, SSO answer owner. Deadline risk: Friday newsletter mention depends on audience, SSO answer and EU exclusion being final by Thursday noon. Post this: "Quick close-the-loop so Friday does not get weird: I have beta email on Thursday, audience = existing trial users only, pricing page pending Legal, demo video pending Leo screenshots, EU customers pending DPA call. Can owners confirm by 10 a.m. tomorrow, especially Legal/DPA/SSO?"
Why this works:
Slack summaries are usually useless because they retell the chat. You need the decisions that survived it. The prompt bans chronology and asks for likely decisions, which catches the slippery middle between rumor and decision.
What to use:
Glean is the cleanest option if it can read the actual channel and adjacent docs. Claude and ChatGPT work if you paste a cleaned dump. Ask for "likely decisions that need confirmation."
📖 AI Alphabet
|
D
|
📖 AI Alphabet Data Drift Data drift happens when the real-world data hitting a model starts to differ from the data it was trained on. When that gap grows, model performance often gets worse. |
AI & Tech News
OpenAI Cuts Expected Microsoft Payments by $97 Billion
The Information reports that revised terms cap OpenAI's payments to Microsoft at $38 billion through 2030, down from a potential $135 billion under earlier targets. The new math matters because the Microsoft-OpenAI relationship is being judged in court and in the market at the same time.
Sutskever Says His OpenAI Stake Is Worth About $7 Billion
Bloomberg reports that Ilya Sutskever testified his OpenAI stake is worth about $7 billion. The number turns the 2023 board fight into a more concrete governance story: mission, control and personal wealth all sit in the same witness record.
Grok Downloads Fall as SpaceX Leases Compute to Anthropic
The Wall Street Journal reports that Grok downloads fell to about 8.3 million in April, down from more than 20 million in January. The awkward backdrop is SpaceX leasing excess compute to Anthropic while Musk's own AI app loses momentum.
Commerce Removes AI Testing Details From Its Website
Reuters reports that the Commerce Department removed public details about its May 5 model-testing agreement with Google, xAI and Microsoft. The deletion puts transparency pressure on a safety process that was supposed to show how major labs would be checked.
Washington Agencies Fight Over AI Evaluation Control
The Washington Post reports that the White House cyber office and Commerce's CAISI are clashing over AI evaluation leadership. The fight shows that model oversight is becoming a power question inside government, not only a policy memo.
npm Attack Hits Mistral, UiPath and TanStack Packages
Socket says a Mini Shai-Hulud supply-chain attack compromised major npm packages, including tools tied to Mistral, UiPath and TanStack's web ecosystem. For builders, the immediate lesson is boring but expensive: dependency audits are now part of AI-era security.
OpenAI Launches Daybreak for Vulnerability Patching
TestingCatalog reports that OpenAI launched Daybreak, a security initiative around Codex Security and faster vulnerability remediation. It lands into the same week as Google's AI-assisted zero-day warning, which makes patch speed the shared theme.
Cognition Hits $445 Million Revenue Run Rate
Colossus reports that Scott Wu's Cognition reached a $445 million revenue run rate in its first 18 months. Devin is no longer only a coding-agent demo story; it is becoming one of the clearest revenue signals in agent software.
GitLab Cuts Jobs to Spend on the Agentic Era
Bloomberg reports that GitLab will cut jobs and reduce its footprint to redirect resources toward growth in the agentic era. Management says the move is not simple AI replacement, but investors will read it as another software company reorganizing around agents.
Robinhood Files for a Second Venture Fund
Axios reports that Robinhood confidentially filed for Robinhood Venture Fund II, targeting early and growth-stage startups. The move extends its push to make private-market exposure feel like a retail product, even as regulators keep watching that boundary.
🚀 AI Profiles: The Companies Defining Tomorrow
Tessera Labs is taking aim at the ERP migration bill nobody likes to sign. The San Jose startup raised $60 million to replace armies of SAP consultants with governed AI agents that move data, code and business logic across old enterprise systems. 🧾
Founders
Founded in 2025 by Kabir Nagrecha and Ming Chang. Nagrecha is the useful kind of prodigy for this market: college at 13, PhD at 20, Meta research experience and a childhood spent near the ERP projects his father ran. Chang brings the co-founder pairing out of pure research and into implementation.
Product
Tessera sells an AI-native system integrator for ERP modernization, starting with SAP ECC to S/4HANA migrations. Its platform breaks transformation work into agent-run steps across data harmonization, legacy-code remediation, cross-system orchestration and audit trails. The promise is not "new ERP." It is fewer humans trapped in the migration basement for three years.
Competition
The obvious opponents are Accenture, Deloitte, IBM, Cognizant and the rest of the consulting machine. Software competitors include Celonis, ServiceNow, SAP's own tooling, Oracle migration partners and every workflow vendor trying to put agents above old systems of record. Tessera's wedge is narrower: own the migration job itself, not another dashboard beside it.
Financing 💰
$60 million Series A led by Andreessen Horowitz, with Foundation Capital, Myriad Venture Partners and Osage University Partners participating. Techmeme's Forbes summary pegged the round at a $320 million valuation; Tech Funding News reported the raise on May 8, 2026. Sources: Techmeme/Forbes, May 7, 2026; Tech Funding News, May 8, 2026.
Future ⭐⭐⭐
ERP migration is a perfect AI market in theory because the work is expensive, repetitive and unpopular. It is also where CFOs discover their "standard" process has 800 exceptions. Tessera wins if agents can produce governed change logs that auditors and business owners trust. Otherwise it becomes a very smart assistant to the same consultants it wants to shrink. 🧱
🤨 Yeah, But...
TechCrunch reported Saturday, citing CNBC, that Nvidia has committed more than $40 billion to equity investments in AI companies in the first months of 2026, including a reported $30 billion OpenAI stake. The same companies are often buyers or prospective buyers of Nvidia hardware, reviving the circular-deal critique.
Our take: Nvidia has solved the classic vendor problem: customers can only buy so many chips with their own money. So the chipmaker is now helping provide the money, which customers can then use to buy the chips, after which everyone can celebrate demand. This is not illegal. It is capitalism wearing a lanyard. The market loves it because each side of the circle has a different label: strategic investment on the way out, infrastructure revenue on the way back. Somewhere in the loop there may be a genuine business, maybe several. But the accounting poetry is doing visible work. Nvidia is not merely selling picks and shovels anymore. It is financing the prospectors, owning part of the mine and invoicing for the digging.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: [email protected]
