AI & Tech Review ⚡

📌 Navigate

📋 Exec Summary

Q2 2022 split AI into two stories: perception and production. DALL-E 2 changed what the public thought AI could do; GitHub Copilot's GA at $10/month changed what enterprises paid AI to do. Google had technically superior image models (Imagen, Parti) and no public product -- a policy gap that would haunt its competitive position through 2023. Stable Diffusion was one quarter away from demolishing the gated-access model entirely.

📊 What Moved

📈 Trend Arcs

Arc 1: Generative Image — Frontier to Commodity in One Quarter

Velocity: Accelerating

April opened with DALL-E 2 behind a waitlist. A handful of researchers and press had access. The images were astonishing and scarce. Scarcity maintained the sense of frontier. By late May, Imagen benchmarks were published (without public access), which that DALL-E 2 was not a singular capability but a reproducible result of combining large diffusion models with CLIP-based conditioning — a recipe that multiple well-resourced teams had independently discovered. By June, the research community understood that text-to-image generation was a solved capability question. The remaining questions were infrastructure (how to serve it cheaply), safety (how to prevent misuse), and access (who gets the weights). Parti's June publication made the point even more forcefully: two different architectures, both from Google, both achieving photorealistic generation, neither available publicly.

The progression matters: in 90 days, text-to-image went from "OpenAI exclusive capability" to "multiple frontier labs have this, the open-source version is one quarter away." The arc accelerated because the underlying technical recipe was sufficiently documented in public research that replication was feasible at the frontier level. Midjourney launched its open beta in mid-April precisely because the capability was reproducible.

Where it stands at quarter close: Three systems — DALL-E 2, Imagen, Parti — represent the frontier, all at different organizations, none fully public. One open-source alternative (Stable Diffusion) is in training. The gating is temporary.

Arc 2: AI-Native Developer Tools — The Productivity Wedge Opens

Velocity: Accelerating

GitHub Copilot had been in technical preview since June 2021. The GA launch on June 21, 2022 was less a capability milestone than a business model one: it established that developers would pay for AI assistance, that $10/month was a price point that cleared psychological resistance, and that the productivity claim was defensible enough to enterprise procurement. The 55.8% faster task completion figure — from a rigorous, pre-registered experiment — gave Copilot a number that competitors could not easily attack. The first-mover advantage in developer tooling is stickier than most categories because workflows become habituated quickly.

The arc's progression across the quarter: April saw continued private beta adoption with growing waitlist demand. May brought enterprise pilot commitments from several large technology companies as procurement teams began evaluating AI tool policies. The June GA launch converted the narrative from "interesting research demo" to "productized service with pricing and SLA." By quarter close, Copilot had >400,000 paying users within days of launch.

What made this arc significant beyond developer tooling: Copilot proved the enterprise AI adoption pattern. The wedge was not autonomous AI (which required trust) but augmentative AI (which required only demonstrated time savings). That pattern repeated across every B2B AI product category over the following two years.

Where it stands at quarter close: Copilot is GA with a paying user base and a pricing model competitors will need to match. The developer tool category is now a proven AI monetization channel. Replit, Amazon CodeWhisperer, and others are accelerating their competitive responses.

Arc 3: The Sentience Question — Noise That Became Signal

Velocity: Accelerating (public discourse); Decelerating (technical credibility)

The LaMDA episode ran from June 11 (Lemoine's Washington Post interview) through late June (his suspension). As a technical claim, it was dismissed within 24 hours. As a cultural event, it was a multi-week story that reached audiences who had never engaged with AI research. The progression: Lemoine's interview was initially treated as a curiosity, then amplified by outlets covering the human-interest angle of an employee challenging Google, then picked up by congressional staffers who began using it as a reference point in AI governance discussions.

The arc's velocity is bifurcated: in technical circles, credibility collapsed immediately and did not recover. In policy and mainstream discourse, it accelerated. By quarter close, questions about AI personhood and moral status had entered legislative briefings in at least three countries. Philosophers who work on consciousness and ethics of mind started receiving unusual quantities of press inquiries.

The important pattern: a demonstrably false technical claim can still inject a true question into public discourse. The question of AI moral status — while not answered by LaMDA's behavior — is a legitimate philosophical question. Lemoine's framing was wrong; the question itself was not. That asymmetry between the discredited claim and the legitimate question it surfaced is what made this arc durable.

Where it stands at quarter close: Lemoine is on administrative leave (terminated July). The consciousness/sentience question is now a live issue in mainstream AI discourse with no clear resolution mechanism. It will compound over subsequent quarters as models become more capable.

🗺️ Landscape Shift

The competitive map at the start of Q2 2022 showed OpenAI in the lead position on generative models with DALL-E 2 as the primary public-facing signal, Google with technically superior models behind an impenetrable policy wall, Microsoft as the primary OpenAI distribution channel via Azure, and no meaningful open-source alternative in the generative image space.

The most important shift in the quarter was not what happened — it was what did not: no major open-source text-to-image model shipped. That absence meant every public-facing image generation product was either gated (DALL-E 2, Imagen) or growing behind a Discord server (Midjourney). The open-source detonation was one quarter away and nobody in the press was modeling it.

💰 Funding & Deal Pattern

Q2 2022 was the inflection point where AI-specific investment started decoupling from the broader venture market contraction. While general VC deployment was pulling back sharply — deal count and dollar volume both fell as rising interest rates and public market corrections forced LPs and GPs to slow deployment — AI deals remained relatively protected.

The capital flow patterns that defined the quarter:

The broader venture context mattered here: Q2 2022 was the quarter when rising interest rates began visibly affecting VC deployment timelines. The free money environment that had supercharged 2021's AI funding was clearly ending.

🔍 The Counter-Narrative

• The consensus: Access-controlled frontier AI (DALL-E 2 waitlists, gated APIs) is the durable distribution model. The reality: Stable Diffusion's open-weight release was 10 weeks away and would let anyone with a consumer GPU run the same class of model. Most investors entering Q3 were positioning for a world of gated access. They were positioning for the wrong world.

• The consensus: DALL-E 2 was the most consequential Q2 launch. The reality: Copilot's GA mattered more structurally. DALL-E 2 changed what people thought AI could do. Copilot changed what enterprises paid for AI to do. Every subsequent AI product that found PMF followed the Copilot template: measurable productivity augmentation, monthly subscription, buyer is a team lead or CTO.

📐 Builder's Benchmark

Capability state at Q2 2022 close:

• Text-to-image: DALL-E 2 accessible via waitlist. No public API pricing. Imagen and Parti benchmark-only, no access. Midjourney Discord beta. Stable Diffusion: not yet available.
• Code generation: GitHub Copilot GA at $10/month individual, $19/month enterprise. Completion-based assistance; no agentic workflows. CodeWhisperer (Amazon): still in preview.
• Language models: GPT-3 (175B) available via OpenAI API at $0.02/1K tokens. Google PaLM not publicly accessible. Meta OPT-175B weights released May 2022 (research use only).
• Compute costs: A100 GPU spot pricing on major clouds was $3-5/hour. Training a DALL-E 2-scale model required millions of dollars of compute — not accessible to most startups.
• API latency: OpenAI API median response time for completions: 2-5 seconds for standard completions. Real-time applications were not viable.
• Open-source gap vs. closed frontier: Significant. The best open-access models (OPT-175B, GPT-J) lagged GPT-3 on most benchmarks. The gap would narrow dramatically in the following 12 months.

Adoption signals:

• GitHub Copilot: >400K paying individual subscribers within the first weeks of GA
• DALL-E 2 waitlist: reportedly 1M+ applicants within weeks of April announcement
• Midjourney: growing to tens of thousands of users in Discord beta
• Stack Overflow's 2022 developer survey (conducted May 2022): 70% of developers using AI tools reported productivity gains

👀 What to Watch

1. Stable Diffusion release (expected August 2022) — The signal is not the release itself but the open-source tooling velocity in the 30 days that follow. If community fine-tuning, ControlNet-style extensions, and web UIs emerge within a month, the gated model strategy is over.

2. OpenAI API pricing changes — GPT-3 at $0.02/1K tokens is still the market price. Any reduction (particularly into sub-$0.01) would signal a shift from scarcity to volume strategy and accelerate application-layer development.

3. Microsoft's enterprise AI rollout pace — Copilot is GA. The question is how fast enterprise procurement moves. Watch for announcements of large enterprise deployments (>10K seats) as the signal that the enterprise sales motion is working.

4. Google's response to the access gap — Google has the best models and no public product. Something has to give — either internal policy changes, a limited external research access program, or a consumer product launch. Any movement on this front before year end would be a significant signal.

5. Funding activity for application-layer AI startups — The pattern in Q2 was infrastructure and developer tools. Q3 will reveal whether capital starts flowing toward vertical AI applications (legal, medical, finance) or whether investors are waiting for the open-source situation to clarify before committing to application-layer bets.

📎 Sources

Key references for this quarter. Links provided where available; historical entries may reference publications by title and date.