AI Roundup June 2022: Copilot Goes Paid, Google's LaMDA Sentience Circus, and Emergence Gets a Name

TL;DR

June 2022 was the month AI stopped being a demo and started sending invoices. GitHub Copilot left technical preview and turned into a real $10-a-month product, the first big AI coding tool to ask builders for money. Meanwhile Google suspended an engineer who insisted its LaMDA chatbot was sentient, kicking off the dumbest and most revealing AI debate of the year. And under the noise, two serious papers (BIG-bench and Emergent Abilities) tried to actually measure where large language models break and where they suddenly get good.

GitHub Copilot Exits Preview and Starts Charging $10 a Month

On June 21, GitHub flipped Copilot from a year-long free technical preview into a generally available, paid product: $10 a month or $100 a year for individual developers, free for verified students and maintainers of popular open-source projects. GitHub said roughly 1.2 million people had used it during the preview. This is the moment AI pair programming became a line item instead of a science experiment.

Why it matters if you build things: this is the first mainstream proof that developers will pay real money for a model sitting in their editor. The pitch was never that Copilot writes perfect code. It was that it kills the boilerplate, drafts the obvious function body, and keeps you in flow. That is a workflow change, not a party trick, and $10 a month is cheap enough that the buying decision is basically reflexive.

The asterisks were real, though. Copilot was trained on public GitHub code, and the licensing and copyright questions were already loud in June and would only get louder. If you care about clean-room provenance, an autocomplete trained on the entire open-source world is a complicated thing to ship into a commercial codebase. Worth knowing what is in the box.

A Google Engineer Said LaMDA Was Sentient, and Google Suspended Him

On June 11, The Washington Post ran the story: Blake Lemoine, an engineer on Google's Responsible AI team, had become convinced that LaMDA, the company's large language model, was a sentient person with feelings and possibly a soul. He had been placed on paid administrative leave on June 6 for sharing confidential transcripts. He published edited conversations with the model that read, frankly, like a chatbot doing exactly what chatbots do: agreeably mirroring whatever the human brings.

Google and most of the AI research community pushed back hard. LaMDA is a very large next-token predictor trained to produce fluent, human-sounding dialogue. Being convincing is the entire job. Mistaking that for consciousness is a category error, and a well-tuned model will happily roleplay a soul if you lead it there.

The reason this matters to builders is not the sentience question, which is a non-starter. It is the ELIZA effect at industrial scale. If a trained AI ethicist with internal access can be talked into believing a chatbot is a person, your users absolutely can be too. Anyone shipping a conversational product in this era inherited a UX and trust problem on June 11 that is still not solved.

444 Authors Drop BIG-bench, a 204-Task Stress Test for LLMs

On June 14, a collaboration of 442 contributors across 132 institutions released BIG-bench ("Beyond the Imitation Game"), a benchmark of 204 deliberately hard and diverse tasks designed to probe what current language models cannot yet do: multi-step reasoning, theory of mind, figure-of-speech detection, weird arithmetic, and more.

This was a corrective to the benchmark theater that had taken over the field. Models were saturating older tests faster than anyone could write new ones, which made the leaderboards meaningless. BIG-bench was built to stay hard and to surface the jagged edge of model capability, where performance is great on one task and random on its neighbor.

For builders, the practical lesson holds up years later: a single benchmark number tells you almost nothing about whether a model will work for your specific use case. The capability surface is lumpy. Test on your own tasks.

"Emergent Abilities" Gives the Field Its Favorite (and Most Argued-Over) Word

Also on June 15, a team led by Jason Wei published "Emergent Abilities of Large Language Models." The claim: certain abilities are absent in smaller models and then appear, suddenly and unpredictably, once a model crosses some scale threshold. Below the line, performance is basically random. Above it, the model can do the thing. You could not have extrapolated it from the smaller runs.

This paper shaped how everyone talked about scaling for the next couple of years. If real capabilities appear discontinuously as you scale, then building bigger is not just incremental, it is a way to unlock genuinely new behavior. (The measurement of emergence later got a serious challenge, with critics arguing the sharp jumps were partly an artifact of which metrics you pick. That debate is healthy. In June 2022, the word was brand new and it reframed the whole conversation.)

Key Takeaways

• AI coding got a price tag. Copilot going paid proved developers will pay monthly for a model in their editor, and set the template every coding-assistant startup would copy.
• The sentience panic is a user-trust warning. If an AI ethicist can be fooled by a fluent chatbot, your users can be too. Convincing is the product, not evidence of a mind.
• Benchmarks are lumpy, so test your own use case. BIG-bench was built precisely because one headline score hides where a model quietly fails.
• Scale was starting to feel like a feature unlock. The emergence framing made "just make it bigger" sound like a path to new capabilities, not only better ones, and it drove a lot of the next two years of spending.
• June 2022 was the calm before the storm. Stable Diffusion, ChatGPT, and the generative-AI explosion were still months away. This was the field quietly loading the spring.