DeepSeek v4, and the end of the OpenAI/Microsoft AGI clause

Simon Willison · email · 2026-04-30 Reader 🖥️

AI Slop AI Unit Economics Evals Inference Optimization Open-Weight Models Trust

A snapshot of late-April 2026 in AI: a Chinese open-weights lab is now pricing the frontier into the floor, the AGI clause that supposedly governed the OpenAI/Microsoft relationship has been quietly written out of existence, and Willison's own LLM tooling has to be rebuilt around the assumption that model output is no longer just text — together describing an industry where the economics, the legal scaffolding, and the developer abstractions are all reorganizing around the same new reality of multi-modal, agentic, commoditized models.

practical · mixed — A weekly newsletter compilation by Simon Willison. Three main expository pieces (DeepSeek V4 release notes, a historical/legal investigation into the OpenAI-Microsoft AGI clause, and a developer log entry about LLM 0.32a0) plus a long tail of curated links, quotes, and notes. The newsletter is primarily practical/observational — reporting what just happened in AI in a single week — but the AGI-clause piece is implicitly argumentative (the clause is dead) and the curated tail surfaces several normative debates (vibe-coding, LLM bans in open source, vegan models).

Structure

Part	Role
DeepSeek V4 release report	Reports the release of DeepSeek-V4-Pro (1.6T) and DeepSeek-V4-Flash (284B) — both MIT-licensed, 1M context MoE models — and quantifies how aggressively they undercut the frontier on price while trailing the state of the art by only 3-6 months
AGI clause history and obituary	Traces the OpenAI/Microsoft AGI clause through OpenAI's own published statements from 2019 through April 27, 2026, ending with the contract restructure that makes Microsoft revenue share 'independent of OpenAI's technology progress' — the clause's de facto death
LLM 0.32a0 refactor write-up	Documents a backwards-compatible refactor of Willison's LLM library: model inputs are now message sequences (not just text), and responses are streams of typed parts (text, reasoning, tool calls, tool outputs, images, audio) instead of a single text stream
Curated tail (links, quotes, notes, releases)	Curates surrounding ecosystem signals: GPT-5.5 prompting, VibeVoice ASR, the 'talkie' vintage 1930 LLM, pip 26.1 lockfiles, Zig's blanket LLM ban, Google Meet speech translation, the 'people do not yearn for automation' essay

Key terms

AGI clause
The provision in the Microsoft-OpenAI agreement that Microsoft's commercial IP rights to OpenAI's technology would terminate once AGI was achieved — with the definition of AGI evolving from the 2018 Charter language ('highly autonomous systems that outperform humans at most economically valuable work') to a $100B-profit-threshold definition (2023 internal docs) to an 'independent expert panel' verification process (October 2025) to functional irrelevance as of April 27, 2026
Mixture of Experts (MoE) Inference Optimization
Architecture where only a subset of model parameters (the 'active' experts) are used per token — DeepSeek-V4-Pro has 1.6T total parameters but only 49B active; Flash has 284B total / 13B active. This is what enables their aggressive pricing.
active parameters AI Unit Economics
The fraction of an MoE model's weights that are actually used to compute each token — directly governs inference cost. Willison emphasizes the active vs total split as the key economic lever.
single-token FLOPs AI Unit Economics
DeepSeek's stated efficiency metric: V4-Pro uses 27% of the FLOPs and 10% of the KV cache of V3.2 at 1M context; V4-Flash uses 10% of FLOPs and 7% of KV cache
streaming parts
Willison's new LLM abstraction: a model response is a stream of typed events (text, reasoning, tool_call_name, tool_call_args, tool_output, image, audio) rather than a stream of text chunks
messages array
The prompt input format adopted by LLM 0.32a0 — a sequence of {role: user|assistant, content} objects, mirroring the OpenAI chat completions API rather than LLM's older 'conversation' object model
pelican riding a bicycle Evals
Willison's signature personal benchmark — asking each new model to produce an SVG of a pelican riding a bicycle. Used as a vibe-check across model generations.
software brain
Borrowed from Nilay Patel: the disposition to model the world as information flows and to see everything as something to be automated; Patel's diagnosis for why AI is unpopular even as ChatGPT usage rises
vegan models Open-Weight Models
Willison's term for LLMs trained entirely on licensed or out-of-copyright data — the 'talkie' 1930 base model qualifies; its chat variant doesn't, because Claude Sonnet/Opus 4.6 were used as judges and synthetic conversation partners during fine-tuning
contributor poker Trust
Loris Cro's framing for Zig's blanket LLM ban: in open source you bet on the contributor, not on the contents of their first PR — reviewing an LLM-written PR yields no future trustworthy contributor, so the ROI is zero
digital smell AI Slop
Andrew Kelley's term for the recognizable stylistic markers of LLM-assisted contributions that abstainers can detect immediately — 'like when a smoker walks into the room'

Propositions

explicitdevelopedDeepSeek-V4-Pro is the new largest open-weights model — 1.6T parameters, larger than Kimi K2.6 (1.1T), GLM-5.1 (754B), and more than twice the size of DeepSeek V3.2 (685B).conf: high
extends ←GLM-5.2 is the new leading open weights model on the Artificial Analysis Intelligence Index · p1 — DeepSeek-V4-Pro is the largest open-weights model (1.6T params) yet places behind GLM-5.2 (744B/40B active) on the Intelligence Index — among open-weight models, raw size does not track general-intelligence leaderboard rank.
explicitcorroboratedDeepSeek-V4-Flash is the cheapest small model on the market, beating even GPT-5.4 Nano; DeepSeek-V4-Pro is the cheapest of the larger frontier models.conf: high
supports ←Howard Yu's Post — American AI models went from 70% of usage to 30% in 12 months · p2 — Corroborates the price premise: DeepSeek-V4 is the cheapest across the market (Flash cheapest small, Pro cheapest of the larger frontier models).
explicitDeepSeek V4's price advantage comes from architectural efficiency gains: V4-Pro uses 27% of the FLOPs and 10% of the KV cache vs V3.2 at 1M context; V4-Flash uses 10% of FLOPs and 7% of KV cache.conf: high
explicitcorroboratedDeepSeek's models trail state-of-the-art frontier models by approximately 3-6 months.conf: high
supports ←Howard Yu's Post — American AI models went from 70% of usage to 30% in 12 months · p5 — Independent confirmation the capability gap is months, not years: Anthropic's 'several months' ahead ~= DeepSeek trailing frontier by ~3-6 months.
explicitThe OpenAI/Microsoft AGI clause — which would have terminated Microsoft's commercial IP rights upon AGI being achieved — is now dead.conf: high
explicitThe definition of AGI in the OpenAI/Microsoft agreement was, as of late 2024, contractually pegged to OpenAI generating $100 billion in profit — not to any technical capability threshold.conf: high
explicitMicrosoft retains a license to OpenAI IP for models and products through 2032, but the license is now non-exclusive; Microsoft revenue share to OpenAI ends; OpenAI revenue share to Microsoft continues through 2030 'independent of OpenAI's technology progress.'conf: high
explicitThe original prompt-in / text-out abstraction LLM was built around in April 2023 is no longer sufficient to represent what frontier models do.conf: high
explicitToday's models return mixed-typed outputs — reasoning, text, tool calls, tool outputs, images, audio — which requires modeling responses as streams of typed parts rather than text chunks.conf: high
explicitPrompts should be modeled as a sequence of typed messages (user/assistant), not as a single text input plus conversation state, because conversations often need to be reconstituted from elsewhere (e.g. building a chat completions API emulator).conf: high
explicitTreating output as typed parts enables better UX: reasoning text can be displayed in a different color on stderr so it doesn't pollute pipelines.conf: high
explicitOpenAI itself now recommends not migrating existing prompts to GPT-5.5 but starting from scratch — 'treat it as a new model family to tune for, not a drop-in replacement.'conf: high
supportsGenerative AI in the Real World: Aishwarya Naresh Reganti on Making AI Work in Production · p14 — OpenAI's own don't-migrate-prompts guidance for GPT-5.5 is direct evidence that model personality differences block portability.
explicitSince GPT-5.4, OpenAI has unified Codex and the main model into a single system — there will be no separate GPT-5.5-Codex.conf: high
explicitAI is unpopular with the general public even as ChatGPT usage skyrockets; the rift is between 'software brain' people (who see the world as something to automate) and everyone else.conf: high
explicitRegular people do not yearn for automation — even decades of effort from Apple/Google/Amazon have failed to make regular people care about smart home automation.conf: high
explicitestablishedA blanket ban on LLM-assisted contributions to an open source project is defensible because OSS reviewing is an investment in the contributor (contributor poker), not just in the code — and reviewing LLM-written code yields no future contributor.conf: high
extends ←Andrew Kelley Interview: Why He Built Zig, Banned AI Contributions, and Moved Zig off GitHub · p3 — This is the primary-source elaboration (fuller quotes, fuller reasoning) of the same 'contributor poker' argument that appears there in brief, secondhand summary form.
supports ←NodeJS Proposes Built-In Virtual File System, Sparking Debate Over AI-Generated Contributions · p4 — Indutny's petition is a real-world instance of exactly the general argument that reviewing LLM-written OSS contributions is a bad investment because it doesn't build a future contributor and can't be adequately verified at scale.
supports ←Who cleans up after the vibe-coding party? · p6 — tldraw's Ruiz banning AI contributions outright is a second named, independent instance of the same maintainer logic behind Zig's blanket LLM-contribution ban -- that OSS review is an investment in a future human contributor, which reviewing AI-authored code cannot repay.
explicitcorroboratedLLM-assisted PRs are detectable in practice — the mistakes LLMs make are categorically different from the mistakes humans make.conf: high
supports ←Andrew Kelley Interview: Why He Built Zig, Banned AI Contributions, and Moved Zig off GitHub · p4 — Kelley's 'we can still tell' laundering anecdote is a concrete illustration of the general claim there that LLM-assisted PRs are detectable in practice.
explicitA 'vegan' base model trained only on out-of-copyright data is feasible (talkie-1930-13b), but currently its chat variant is impure because modern LLMs are used as judges and synthetic conversation partners during fine-tuning.conf: high
explicitA consumer Mac (128GB M5 MacBook Pro) can now plausibly run quantized versions of frontier-adjacent models locally — Willison expects DeepSeek-V4-Flash to fit, and is hopeful even Pro could stream active experts from disk.conf: medium
explicitMatthew Yglesias's vibecoding sentiment: he doesn't want to vibecode — he wants professionally managed software companies to use AI coding assistance to make better, cheaper software products that they sell to him.conf: high

Arguments

The OpenAI/Microsoft AGI clause is dead. strength: high

Explicit premises:

OpenAI's 2019 launch language committed to licensing only 'pre-AGI' technologies to Microsoft
By late 2024, AGI in the contract was operationalized as a $100B profit milestone — not a capability threshold
October 2025 introduced an 'independent expert panel' to verify AGI declarations
February 2026 OpenAI publicly affirmed 'AGI definition and processes are unchanged'
April 27, 2026 contract restructure makes OpenAI's revenue share to Microsoft 'independent of OpenAI's technology progress' through 2030, with Microsoft's license non-exclusive through 2032
The Verge reached the same conclusion in their reporting

Implicit premises:

If revenue share is explicitly decoupled from 'technology progress' (the only mechanism by which AGI achievement would change anything), the AGI trigger has no remaining contractual function
OpenAI's silence on the AGI clause itself in the April 27 statement, juxtaposed with their February affirmation that it was 'unchanged,' is the diplomatic way of retiring it

Inferences:

AGI as a contractual concept never survived contact with commercial reality — every attempted definition (Charter language, profit threshold, expert panel) reveals more about the corporate dynamics than about AGI itself
Future commercial agreements pegged to AGI achievement should be treated skeptically — the precedent now is that such clauses get renegotiated out before they can ever fire

DeepSeek V4 brings frontier-adjacent capability at a fraction of the price of US frontier models. strength: high

Explicit premises:

MIT-licensed, 1M context, MoE architecture: Pro 1.6T/49B active, Flash 284B/13B active
Pricing: Flash at $0.14/$0.28 per million tokens input/output; Pro at $1.74/$3.48 — undercuts GPT-5.4 Nano and other frontier models
Architectural efficiency gains documented in DeepSeek's paper (FLOPs, KV cache reductions)
Self-reported benchmarks place V4-Pro within 3-6 months of state-of-the-art

Implicit premises:

Pricing reflects underlying inference cost (not subsidy) — the FLOP reductions support this
Open weights plus MIT license means the price floor will fall further once Unsloth and others release quantized variants for local inference
Trailing the frontier by 3-6 months is acceptable for the majority of use cases

Inferences:

The cost moat that justified frontier-lab pricing is shrinking rapidly; the differentiator must now be something other than raw model quality (UX, tooling, agentic capability, distribution)
Local/private inference becomes more viable for serious workloads as consumer hardware approaches the threshold to run quantized Pro models

LLM tooling must shift from prompt/response to messages/typed-parts because model outputs no longer fit the text-in / text-out shape. strength: high

Explicit premises:

The 2023 LLM library was built for text in / text out
Models now accept multi-turn message arrays, multimodal attachments, schemas, and tools
Models now return mixed output: reasoning, text, tool calls, tool outputs, images, audio — and may execute server-side tools (code interpreter, web search)
The old conversation object made it hard to feed in pre-existing conversations (e.g. for chat-completions API emulation)

Implicit premises:

An abstraction should faithfully represent what the underlying systems actually do — leaky abstractions force workarounds (e.g. SQLite-only conversation reinflation)
Future model capabilities (more modalities, more typed outputs) are easier to add to a typed-parts stream than to a text stream

Inferences:

Existing prompt-engineering tools and wrappers built on the prompt/response model will face similar refactors as multimodal/agentic capabilities mature
Serializing/deserializing responses (response.to_dict / from_dict) becomes a core feature, since these richer objects can't be reconstructed from a string

A blanket ban on LLM-assisted contributions is rational for projects whose primary review goal is cultivating contributors, not landing code. strength: high

Explicit premises:

In successful OSS projects, PR review is an investment in the contributor's growth
LLMs let a stranger submit a perfect PR that teaches the project nothing about the human behind it
Reviewer time is finite — reviewing an LLM PR yields no future contributor
If a PR was mostly LLM-written, the maintainer could just fire up their own LLM to solve the problem

Implicit premises:

Trust in OSS is accumulated through humans demonstrating judgment over time — LLM-mediated submissions short-circuit that signal
Andrew Kelley's claim that LLM 'digital smell' is detectable makes enforcement viable in practice

Inferences:

Projects whose review purpose is purely 'land good code' (e.g. monorepo with paid maintainers) have less reason to ban LLM contributions than community-driven OSS
There is no universal answer — the policy follows from the project's theory of value

Narratives

The AGI clause's seven-year evolution: From the 2018 OpenAI Charter ('highly autonomous systems that outperform humans at most economically valuable work'), through the 2019 Microsoft partnership announcement promising to license only 'pre-AGI technologies,' to the 2024 disclosure that AGI was internally defined as the systems capable of generating $100B in cumulative profit, to the October 2025 introduction of an 'independent expert panel,' to OpenAI's February 27, 2026 affirmation that 'AGI definition and processes are unchanged,' to the April 27, 2026 restructure that explicitly decouples Microsoft's revenue share from 'OpenAI's technology progress.'
→ The chronological structure is the argument — each successive redefinition shows AGI being domesticated from a civilizational threshold into a balance-sheet milestone into a panel decision into an irrelevance
Matt Levine's 2023 hypothetical (Willison's 'all-time favorite' commentary): OpenAI achieves benign superintelligence, cuts its investors a check for their capped return, and solves all of humanity's problems — capitalism ends, and nobody needs Microsoft products.
→ Crystallizes why the AGI clause was always a paradox: any AGI capable of triggering the clause would also obsolete the commercial structures the clause was meant to govern. Reinforces the obituary tone.
The LLM library's evolution from April 2023 to April 2026: started as a wrapper around text-in/text-out OpenAI calls, accumulated attachments (image/audio/video), then schemas (structured JSON output), then tools (tool calls), then reasoning support, then multimodal output — until the original abstraction had to be retired for a typed-parts stream.
→ Personal-developer-history-as-microcosm: shows in miniature how the AI industry's last three years have stretched every abstraction past its breaking point.
The pelican-on-a-bicycle benchmark: Willison's running test where every new model generates an SVG of a pelican riding a bicycle. DeepSeek-V4-Flash produces a respectable pelican with mean expression; DeepSeek-V4-Pro produces a one-winged hairy-backsided pelican that looks 'drawn by a different artist from the bicycle.' Scott Jla's escalation: a horse riding an astronaut riding a pelican riding a bicycle, with ChatGPT Images 2.0 spontaneously adding a road sign 'WHY ARE YOU LIKE THIS.'
→ Vibe-check for model generations; signals that improvements come in unexpected and uneven ways; offers a personality counterweight to the otherwise transactional release reports
talkie-1930: Nick Levine, David Duvenaud, and Alec Radford release a 13B 'vintage' language model trained on 260B tokens of pre-1931 English text. When asked to generate an SVG of a pelican on a bicycle, it refuses to acknowledge the request and instead asserts that 'An SVG of a pelican riding a bicycle was generated in 1860... The idea was suggested by the fact that pelicans are sometimes seen fishing on horseback, on the banks of the Rhine.'
→ Demonstrates that 'vegan' models trained on pre-cutoff data behave fundamentally differently — they hallucinate in period style; their training data shapes not just facts but mode of expression. Charming and slightly unsettling.
Bun's 4x Zig compile improvement: Anthropic-acquired Bun built parallel semantic analysis and multiple codegen units in their Zig fork — and explicitly will not upstream it to Zig because Zig has banned LLM-authored contributions.
→ Concrete case study of what the Zig LLM ban actually costs the broader Zig ecosystem — a real performance improvement remains stuck downstream. Frames the policy debate with stakes.

Questions addressed

✅ What is DeepSeek V4 and how does it compare to existing frontier models?
Open-weights MoE in two sizes; trails frontier by 3-6 months; undercuts everyone on price; can plausibly run locally on high-end consumer hardware in quantized form
✅ Why can DeepSeek price so aggressively?
Architectural efficiency — order-of-magnitude reductions in single-token FLOPs and KV cache at long context, especially in Flash
✅ How was AGI defined in the OpenAI/Microsoft contract, and what happened to that definition?
It evolved from the Charter language to a $100B profit threshold to an expert-panel verification process, and is now functionally retired as of April 27, 2026
✅ Why does LLM 0.32 need such a big refactor?
Because model I/O has fundamentally changed — multi-turn message arrays as input, typed-parts streams (text, reasoning, tool calls, tool outputs, multimodal) as output
✅ Why is the general public hostile to AI even as ChatGPT usage grows?
Per Nilay Patel: a 'software brain' worldview that treats everything as automatable has alienated everyone whose work and relationships resist database modeling
✅ Should open source projects ban LLM-assisted contributions?
Depends on the project's theory of value — for projects investing in contributor growth (contributor poker), a blanket ban is defensible; the practical detectability of LLM contributions makes enforcement viable
⭕ Is it possible to train an LLM exclusively on out-of-copyright / licensed data?
Base models yes (talkie-1930-13b); but useful chat variants currently require modern LLMs as judges and synthetic conversation partners, contaminating the purity. The talkie team aspires to bootstrap entirely from vintage data eventually.
⭕ What is the strategic implication for OpenAI/Microsoft of the AGI clause being retired?
Willison reports the facts but does not speculate. The implication — that AGI rhetoric is decoupling from contractual reality, and that the partnership is normalizing into a standard non-exclusive license — is left to the reader.

Gaps

incompleteThe DeepSeek piece reports prices and FLOP reductions but does not address the broader geopolitical / supply-chain context — US export controls, the H100/H800 distinction, the question of whether DeepSeek's efficiency gains are partly a forced response to compute constraints rather than pure research progress. The framing is purely technical/commercial.
Read next: ChinaTalk / Jordan Schneider on US-China AI competition; Dwarkesh Patel interviews with DeepSeek-adjacent researchers; SemiAnalysis on Chinese AI inference economics
incompleteThe AGI clause obituary documents the legal mechanics but does not interrogate what the original clause was actually for. Was it ever a real safety mechanism (the Charter framing) or always a negotiating chip / PR talking point? The piece presents the dissolution as inevitable without examining whether anyone inside OpenAI ever took the clause seriously as governance.
Read next: Karen Hao's 'Empire of AI'; Dario Amodei's departure rationale; Ilya Sutskever's Superalignment exit; the actual text of the OpenAI Charter and its drafting history
incompleteThe LLM 0.32 refactor describes the technical changes but does not address how downstream plugins and users will be affected, beyond noting that previous prompt= calls still work. The migration story for the broader ecosystem (every llm-* plugin, every script that does for chunk in response: print(chunk)) is implicit.
Read next: The LLM plugin documentation; comparison with how Anthropic, OpenAI, and Google SDKs have evolved their streaming APIs
incompleteThe Zig LLM ban discussion presents the maintainer-perspective argument well (contributor poker) but does not engage with the contributor-perspective question: if LLM assistance lowers the barrier for new contributors with imperfect English or self-taught backgrounds, does a blanket ban systematically exclude marginalized contributors?
Read next: Research on barriers to OSS contribution (Bogdan Vasilescu et al.); Drew DeVault's writing on Zig and OSS sustainability; debates around the EU AI Act and OSS
closed byWho cleans up after the vibe-coding party? — Karen Sandler's interview supplies the missing contributor-perspective counter-argument to blanket AI-contribution bans that the DeepSeek/Zig gap flags as absent: that shunning these tools could make free software fall behind proprietary software, and that they carry real democratizing potential to let more people modify the software they depend on. This answers the gap's general question (is there a legitimate contributor-side case against bans?) though not its specific framing around marginalized/non-native-English contributors.
unaddressedWillison's newsletter style is to compile signals and let connections emerge implicitly. The three lead stories share a deep unity — commoditization (DeepSeek), legal/conceptual dissolution (AGI clause), and abstraction collapse (LLM refactor) all point at the same phenomenon of AI no longer being the bounded, specialized, frontier-lab artifact it was in 2023 — but the newsletter never makes this synthesis explicit.
Read next: Ben Thompson's Stratechery on disaggregation; Carlota Perez on technological surges and the transition from installation to deployment phase; Simon Willison's own blog post 'Open Weights = Open Doors' if it exists, or his back catalog on commoditization patterns

DeepSeek v4, and the end of the OpenAI/Microsoft AGI clause

Structure

Key terms

Propositions

Arguments

Narratives

Questions addressed

Gaps

Related documents