AI & Tech Review ⚡

📌 Navigate

📋 Exec Summary

Q1 2022 rewired AI's competitive logic around two pivots: InstructGPT proved alignment is a capability multiplier (not a constraint), and Chinchilla proved compute-optimal training beats brute-force scaling. Together they collapsed the "bigger model wins" thesis and set the agenda for every training run through 2024. Meanwhile, text-to-image generation reached commercial threshold with Midjourney's closed beta, and OpenAI maintained the only frontier model API on the market.

📊 What Moved

📈 Trend Arcs

Arc 1: Alignment Research Enters the Product Layer

Velocity: Accelerating

January 2022 was the moment alignment research crossed from academic safety concern to product differentiator. Prior to InstructGPT, the dominant assumption was that RLHF and related preference-learning techniques were safety mechanisms — guardrails on dangerous outputs, not tools for making models more useful. The InstructGPT paper inverted this framing by demonstrating that a 1.3B parameter model trained with human feedback could produce outputs that human evaluators preferred over GPT-3 outputs at 175B parameters on open-ended generation, QA, and instruction following. The finding meant alignment was not in tension with capability — it was the mechanism through which capability became legible to users.

Throughout Q1, the research community absorbed this result and began extending it. Papers from Anthropic's early work, DeepMind's work on reward modeling, and independent researchers on the safety-capabilities-overlap expanded the surface area of RLHF as a technique. By March, the major AI labs had either adopted RLHF as a core training component or were actively experimenting with it. The alignment-product convergence that defined 2023's assistant race was locked in during Q1 2022, nine months before the product that made it visible to the public.

Where it stands at quarter close: InstructGPT is deployed on OpenAI's API as the default completion model. Anthropic has incorporated RLHF into its core research program. The technique is no longer experimental — it is the new baseline for fine-tuning frontier language models.

Arc 2: Compute-Optimal Training Replaces Brute-Force Scaling

Velocity: Accelerating (from near zero)

Through 2020 and 2021, the AI field's dominant strategy was the one the Kaplan scaling laws seemed to endorse: train bigger models. The arms race had a clear scoreboard — parameter count — and every major lab competed on it. GPT-3 at 175 billion parameters (2020) was followed by Gopher at 280 billion (December 2021), Megatron-Turing NLG at 530 billion (January 2022), and PaLM at 540 billion (announced April 2022, largely trained through Q1). The assumption embedded in all of this was that more parameters meant better performance and that compute was best spent on model size rather than data volume.

Chinchilla broke the arc in March. The paper was precise: given a fixed compute budget, you should train a model roughly equal in parameter count to the number of training tokens (loosely — the exact formula is budget-dependent). For a model the size of Gopher, compute-optimal training would have used 1.4 trillion tokens instead of 300 billion. The conclusion: Gopher was undertrained by a factor of nearly 5x on data relative to its compute budget. Every model trained under the Kaplan assumptions was similarly miscalibrated.

The velocity shift was structural. Labs could not easily rerun their flagship models under the new regime mid-quarter. But the paper set the agenda for everything trained after March 2022. LLaMA (February 2023) was explicitly Chinchilla-scaled. Mistral's series (2023) was Chinchilla-scaled. The open-weight movement's viability — competitive quality at small model sizes — depended on a scaling law published in Q1 2022.

Where it stands at quarter close: The Chinchilla paper is public and widely circulated in the research community. No production models have yet been retrained under its recommendations. The impact will be felt in Q2–Q4 2022 as new training runs are designed.

Arc 3: Text-to-Image Generation Reaches Commercial Readiness

Velocity: Accelerating

DALL-E (the original, January 2021) demonstrated that text-to-image generation was possible at scale. Throughout 2021, diffusion models — GLIDE, CLIP-guided diffusion, and related work — improved on the GAN-based approaches that had dominated generative image research. By Q1 2022, the capability curves in the research literature were pointing toward commercial-quality generation within quarters, not years.

The key developments through Q1 2022 were less about headline releases and more about the accumulation of techniques: classifier-free guidance improving sample quality dramatically, CLIP embeddings providing the text-image alignment mechanism, and cascaded diffusion architectures enabling high-resolution outputs. DALL-E 2's April release drew directly on this stack — most of the underlying components were public research by March 31, 2022.

The practical signal for builders: if you were reading arXiv through Q1 2022, the arrival of commercial text-to-image generation was not a surprise. It was a predicted outcome from a research pipeline whose components were visible. The builders who got early access to DALL-E 2, Midjourney (launched March 2022 in closed beta), and Stable Diffusion (August 2022) had their first-mover advantage because they were reading primary sources rather than waiting for product launches.

Midjourney quietly launched its closed Discord beta in March 2022. This was the first commercial text-to-image product accessible to non-researchers. It was rough, slow, and hard to use. It also had a waitlist that filled immediately.

Where it stands at quarter close: Midjourney is in closed beta. DALL-E 2 is days away from announcement. Stable Diffusion is in training. The text-to-image arc has not yet crossed into mainstream visibility but is at commercial threshold.

🗺️ Landscape Shift

The competitive map in Q1 2022 was still organized around a simple axis: who had the biggest model. By March 31, the Chinchilla paper had introduced a second axis — compute efficiency — that would reorganize the field over the following 18 months.

The structural shift: OpenAI expanded its moat not by releasing a bigger model but by releasing a better-aligned smaller model. DeepMind raised its research credibility without a product deployment. Anthropic remained in the research phase. The absence of a strong third commercial API competitor to OpenAI was the defining structural fact of Q1 2022 — a gap that would not close until mid-2023.

💰 Funding & Deal Pattern

Q1 2022 AI investment remained elevated but showed early signs of the correction that would intensify through 2022. The 2021 SPAC-fueled environment had begun cooling in public markets by Q4 2021; that cooling reached private venture with a lag.

🔍 The Counter-Narrative

• The consensus: Bigger models win; parameter count is the scoreboard. The reality: Three consecutive quarters of evidence (FLAN, InstructGPT, Chinchilla) showed smaller, better-trained, better-aligned models beat bigger raw models. The builders who internalized this were reading primary sources while everyone else read press releases.

• The consensus: Open-source AI is years behind frontier. The reality: The capability gap was real, but the enabling conditions for parity were already in motion — EleutherAI proved open training at scale was feasible, Chinchilla made it economically viable, and Meta's OPT and Stable Diffusion were in development. The gap's closure was scheduled, not speculative.

📐 Builder's Benchmark

API pricing (Q1 2022 reference points):

• OpenAI GPT-3 (Davinci): $0.02 per 1K tokens (approximately $20 per million tokens)
• OpenAI Codex: Free during beta (launched August 2021, still in beta access)
• No other frontier model APIs available commercially
• Token pricing context: a typical enterprise use case (document summarization, classification) at scale would cost $2,000–$10,000 per month at Davinci pricing; the cost made many applications economically marginal

Performance benchmarks that shifted this quarter:

• InstructGPT 1.3B outperforms GPT-3 175B on human preference evaluation for instruction-following tasks: a 130x parameter reduction with positive preference outcome
• AlphaCode (DeepMind, released January 2022): reaches roughly 50th percentile on competitive programming benchmarks — first model to demonstrate competitive-level code generation
• Chinchilla (70B) outperforms Gopher (280B) on MMLU, BIG-bench, and the majority of NLP benchmarks evaluated; parameter efficiency as a benchmark axis introduced formally

Adoption curves:

• OpenAI API developer accounts: estimated 1M+ by Q1 2022 close (not publicly disclosed; inferred from product blog references)
• GitHub Copilot: still in technical preview; not yet generally available (GA launched June 2022)
• Midjourney Discord: closed beta launched March 2022; waitlist filled within days

Open-source vs. closed gap:

• Capability gap: large. Best open weights (EleutherAI GPT-NeoX-20B) vs. GPT-3.5 equivalent: significant quality difference on most tasks
• Infrastructure gap: narrowing. Open training tooling (Megatron-DeepSpeed, accelerate library) improving rapidly
• Economic gap: closing. Chinchilla paper published March 2022 — open-weight teams now have the scaling law that makes smaller, cheaper training competitive

👀 What to Watch

📎 Sources

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