Interaction Models Unlock Continuous Human-AI Collaboration

Enterprise AI's Support Imperative

Powerful AI models alone cannot overcome institutional inertia in enterprises; dedicated forward-deployed engineering is essential for deep transformation. OpenAI's new joint venture, DeployCo, secures $4B initial investment at $10B pre-money valuation from partners like TPG, Advent International, Bane Capital, Brookfield, and Goldman Sachs. Structured as a separate company, it acquires engineering firm Tomorrow for 150 staff experienced in AI deployments, prioritizing access for investors' portfolio companies. This mirrors Anthropic's unnamed effort, signaling a wave of M&A to meet surging demand. Smaller agencies serve the long tail, but massive tonnage of onboarding support will be needed for decades.

Discourse reaffirms that models crash into organizational hurdles without external structures. Investors buy priority access, underscoring the capability overhang enterprises must close through human-led integration.

Private Market Perils in AI Stocks

Secondary markets for private AI stocks, often via blockchain tokens backed by SPVs, face voiding by labs. Anthropic updated support docs to declare unapproved transfers void, listing firms offering public sales as fraudulent or valueless due to transfer restrictions. OpenAI echoed this in a blog post, stating unauthorized transfers are legally void. Lawyer Gabriel Shapiro warns of lawsuits against Anthropic and platforms like Forge; prices crashed 50% post-notice.

This reflects broader private market dynamics post-GFC: delayed IPOs, unlimited private capital, and synthetic ownership via layered SPVs (e.g., tokenized Cayman/Delaware vehicles). Retail/accredited investors seek exposure but lack verification or SEC protections. Critics like Casey Craig call positions 'four layers from actual shares.' Potential reckoning looms if labs invalidate structures, especially post-lockups, though dozens of registered SPVs limit aggressive voiding.

Policy Shifts: Light-Touch Safety and Trade Signals

White House walks back FDA-like AI safety regime after NEC Chairman Kevin Hassett's offhand comparison sparked backlash. Hassett clarified no new bureaucracy; instead, direct collaboration with labs ensures models avoid extreme harm pre-release. Former AISAR David Sacks confirmed no senior support for heavy regulation.

Trump assembles tech envoy for China trip: Elon Musk, Tim Cook, Dina Powell McCormack, plus finance/semiconductor execs (Micron, Qualcomm; notably absent: Jensen Huang). Aims to finalize trade negotiations, bilateral board; stalled H200 GPU exports signal AI chips off-table.

Redefining Interaction: From Turn-Based to Continuous

Current AI mimics email—discrete turns freeze perception, batching user thoughts, blocking pointing/clarifying/interrupting. Thinking Machines, founded by ex-OpenAI CTO Mira Murati with superteam (incl. Barrett Zoph, Luke Metz who returned to OpenAI; John Schulman), counters with interaction models trained from scratch for continuous time-aware exchange.

Architecture splits streams into 200ms micro-turns: real-time interaction model handles parallel audio-visual input/output (perceiving/responding simultaneously); background model manages reasoning/tools/agents. Enables multitasking: e.g., chatting while searching (Devil Wears Prada 2 box office query pulls real-time data outside training cut-off).

Demos showcase: new person detection; real-time translation (phrase-by-phrase); dialogue management (track thinking/yielding); visual interjection (slouch reminder); professional softening (rephrase blunt colleague message); proactive speech (timed breathing, code-switch translation). New benchmarks: TimeSpeak (initiate speech at specified times with correct content); QSpeak (semantically apt timing/responses).

Mira Murati: "The current AI experience often feels like a conversation that only begins after we stop talking... Interactivity has to be in the model and it has to scale with intelligence."

This prioritizes human-AI bandwidth over raw autonomy, raising combined intelligence ceiling while keeping humans central. Sumit Chentelal: "Secret plan: 1. Increase human-to-AI bandwidth 2. Raise ceiling of human+AI intelligence 3. Help humans continue as main characters."

Rowan Zeers: First general video+speech model that's visually proactive. John Schulman emphasizes underemphasized human-AI collaboration capabilities.

Key Takeaways

• Enterprises need forward-deployed engineers like DeployCo to bridge AI capability overhang and inertia; expect M&A surge.
• Avoid secondary AI stock tokens/SPVs—labs declare them void, risking zero value and lawsuits.
• US AI policy favors lab collaboration over FDA-style bureaucracy for pre-release safety checks.
• Trade talks signal Nvidia AI chips excluded from China deals amid export stalls.
• Interaction models process 200ms micro-turns for continuous collaboration, enabling real-time translation, proactivity, and background tasks.
• Build AI around human workflows: real-time model + background reasoning boosts 'situational smartness.'
• Invent benchmarks like TimeSpeak/QSpeak for proactive audio capabilities.
• Prioritize interactivity scaling with intelligence to maximize human+AI potential.
• Thinking Machines advances non-frontier labs via focused human-AI paradigms.