Top 10 AI News and Developments: April 23 - April 30, 2026

Executive Summary

The week's signal density was unusually high, but the deeper coherence is structural rather than incremental. Three independent threads converged: a frontier base-model release that re-pretrains for the first time in two generations, a billion-dollar realignment of the cloud-AI alliance graph, and an open-weight model from China that closes the proprietary gap on coding while running on a single consumer GPU. The frontier story is OpenAI's GPT-5.5 "Spud", the first fully retrained base model since GPT-4.5, posting 82.7% on Terminal-Bench 2.0 with roughly 40% fewer output tokens for equivalent Codex tasks (Handy AI Substack, O-Mega). The open-weight story is DeepSeek V4-Pro and V4-Flash, a 1.6T-parameter MoE shipped on April 24 after a year-long detour through failed Huawei Ascend training and a return to NVIDIA hardware. The capital-and-cloud story is twofold: Microsoft and OpenAI tore up their exclusivity arrangement on April 27, and within 24 hours OpenAI's frontier models, Codex, and a new Bedrock Managed Agents service appeared on AWS, while Google committed up to $40B to Anthropic on top of Amazon's $25B from days earlier.

These stories interlock. The Microsoft-OpenAI restructure removes the AGI clause and the cloud-exclusivity that have anchored the industry's commercial assumptions for six years; it explicitly enables OpenAI's Bedrock-Managed-Agents launch the very next day; it gives Microsoft optionality to lean harder into Anthropic, which is now bilaterally funded by Google ($10B immediate, $30B milestone-gated, plus 5GW of compute starting 2027) and Amazon ($25B). The hyperscaler-AI lab graph has gone from triangular and exclusive to fully meshed and non-exclusive. At the same time, the open-weight tier is no longer chasing the frontier from behind. Qwen3.6-27B — a dense 27B model with a Gated DeltaNet hybrid architecture — beats Alibaba's own 397B MoE flagship on agentic coding benchmarks while quantizing to 16.8 GB and running at 25 tokens/sec on consumer hardware, and DeepSeek V4-Pro's tech report concedes only a "three-to-six-month" gap to GPT-5.4 and Gemini 3.1 Pro.

The architecture and labor stories complete the frame. Two AI-in-science papers landed: an Emory PNAS paper used a structurally constrained neural network to discover new non-reciprocal force laws in dusty plasma at 99% accuracy, and a multi-institution arXiv paper titled "There Will Be a Scientific Theory of Deep Learning" argues that the field is mid-transition from alchemy to a predictive science of "learning mechanics." On the labor side, Meta announced 8,000 layoffs and 6,000 unfilled roles, and Microsoft offered voluntary buyouts to ~7% of its US workforce, with both companies citing AI-driven productivity reallocation. And the Anthropic Mythos cybersecurity model is under investigation for unauthorized access via a third-party vendor — the first material-incident-disclosure on a deliberately gated frontier model. Apple's confirmed pivot to a custom 1.2T-parameter Gemini variant for the next Siri closes the loop: even the most vertically-integrated consumer-AI player is now sourcing foundation models from a competitor.

The unifying read is that the assumptions of 2024-2025 — exclusive Microsoft-OpenAI, vertically integrated Apple AI, frontier-only-from-three-US-labs, transformer-only architectures — are all weaker than they were a week ago. What replaces them is an industry that looks less like an oligopoly and more like a multi-party meshed graph with much more capital, much more non-attention architecture in production-adjacent code, and a near-frontier open-weight tier that runs on consumer hardware.

1. OpenAI Ships GPT-5.5 "Spud" as First Full Pretraining Retrain in Two Generations

OpenAI released GPT-5.5 on April 23, 2026, the first fully retrained base model since GPT-4.5; every prior GPT-5.x release (5.0 through 5.4) was a post-training iteration on the same foundation (O-Mega complete guide). The codename "Spud" had been kicking around partner decks for months as a putative GPT-6, but OpenAI shipped it under the 5.5 banner — a signal that incremental versioning is now the deliberate strategy and that the next major version bump will be reserved for whatever combination of architectural and capability changes the company considers truly discontinuous (Wikipedia entry).

The benchmark headline is Terminal-Bench 2.0 at 82.7%, a decisive lead over Claude Opus 4.7's 69.4% and Gemini 3.1 Pro's 68.5%, and FrontierMath performance of 51.7% on tiers 1-3 and 35.4% on tier 4 (Handy AI Substack). Notably, GPT-5.5 trails Opus 4.7 on SWE-Bench Pro (58.6% vs Opus 4.7's 64.3%), making this the first frontier release in months where the benchmark surface clearly differentiates between coding-task subtypes. The interesting operational claim is "~40% fewer output tokens than GPT-5.4 for equivalent Codex tasks" at matching per-token latency — a reasoning-density gain that translates directly into agent step economics and is more important than the raw benchmarks for any production deployment running long agent horizons. The model ships in three consumer surfaces (default, Thinking, Pro) with five reasoning-effort levels (xhigh through non-reasoning), 1M-token API input context, 128K output, and pricing of $5 / $30 per million input/output tokens for the standard tier and $30 / $180 for Pro — a 2x bump on standard input and unchanged for Pro relative to GPT-5.4.

The strategic content is the retrain itself. OpenAI is signaling that post-training on the GPT-4.5 foundation has hit a meaningful diminishing-returns curve, and that the future capability frontier requires base-model resets. Pre-training reportedly completed on March 17 with post-training in early April, and the slip from the originally telegraphed April 14 ship date is widely attributed to safety/alignment review rather than capability gating (Fazm article). For an industry that had begun to treat frontier-model architecture as effectively static, the Spud retrain is a reminder that the foundation layer is still under active revision, even if the visible cadence of updates suggests otherwise.

2. DeepSeek Ships V4-Pro and V4-Flash After Year-Long Hardware Detour

DeepSeek released V4-Pro and V4-Flash on April 24, 2026, after a 16-month gap from V3 that included a failed training run on Huawei Ascend 910B silicon and a forced return to NVIDIA H20 hardware (Fortune coverage, CNBC). V4-Pro is a 1.6 trillion-parameter Mixture-of-Experts model — the largest open-weight model ever released, surpassing the prior record holder by a meaningful margin — with a 1M token context window. V4-Flash is the cheaper, smaller variant aimed at high-throughput agentic deployments. Both are released under a permissive open-weight license that preserves DeepSeek's pattern from V3.

The performance claim that matters for the field is in DeepSeek's own tech report: V4-Pro "falls marginally short of GPT-5.4 and Gemini 3.1 Pro," with the company explicitly framing the result as "a developmental trajectory that trails state-of-the-art frontier models by approximately three to six months." That the leading Chinese lab is publishing self-assessed proximity-to-frontier with this kind of precision — and that the proximity has tightened from "roughly a year" in the V3 era to "three to six months" now — is the more important signal than the parameter count. Against open-source competition, V4-Pro is reported as the leader on agentic coding and reasoning benchmarks, displacing GLM-5.1 and Kimi K2.6 in the upper tier of openly-licensed frontier-class systems.

The geopolitical and operational subtext is the Huawei detour. V4 was originally targeted for Q3 2025 on domestic silicon; the Ascend 910B training failures forced a roll-back to NVIDIA H20 (which itself sits inside US export-control constraints), with the lite 200B variant released on March 9 as evidence that the architecture was finalized while capacity for the full 1T+ run was being secured (OSINT report on Reddit). The lesson is that domestic Chinese silicon has not yet hit the maturity required for trillion-parameter frontier training, but the architectural and post-training pipeline can absorb a hardware migration without losing the underlying model. For practitioners, the V4-Pro/Flash pair is now the strongest open-weight option for tasks that require frontier reasoning at zero per-token API cost.

3. Microsoft and OpenAI Tear Up the AGI Clause and the Cloud Exclusivity

Microsoft and OpenAI announced a renegotiated long-term partnership on April 27, 2026, that eliminates the cloud exclusivity, removes the AGI-trigger clause, and restructures the revenue-sharing relationship in both directions (Computerworld, CNBC, NYT). Under the new terms, Microsoft retains a non-exclusive license to OpenAI's IP through 2032, Azure remains the "primary" cloud and gets first-ship rights for new products "unless Microsoft cannot and chooses not to support the necessary capabilities," but OpenAI can now distribute "all of its products" across any cloud provider. Microsoft will no longer pay OpenAI a revenue share; OpenAI continues paying Microsoft a 20% revenue share through 2030, but now subject to a total cap.

The AGI clause removal is structurally the most important change. In the prior agreement, Microsoft's exclusive access to OpenAI's IP would lapse upon achievement of artificial general intelligence — a definition both companies have spent the last two years quietly arguing about, with revenue-share-cap implications running into the trillions of dollars on the upside. By removing the clause entirely and replacing it with a fixed 2032 license expiration plus a revenue cap, both parties have converted a contingent unbounded liability into a bounded contractual structure. The trade is that Microsoft loses the optionality on AGI-driven exclusive access, and OpenAI loses the optionality on triggering an exit from the revenue share by declaring AGI. Each side is buying certainty.

The downstream implications are mechanical and immediate. OpenAI now ships across all clouds, which manifested less than 24 hours later as the Bedrock launch (Story 4). Microsoft is now contractually free to "lean harder into its own models, into Anthropic, and into whatever else the market produces," in the words of analysts cited in the Computerworld coverage. The corporate-AI market structure shifts from a triangular Microsoft-OpenAI-and-Anthropic-via-Amazon arrangement to a fully meshed graph. For enterprise IT, the practical change is that "OpenAI on Azure" is no longer a privileged deployment path — the same models with the same licensing terms now run on AWS, and Microsoft's competitive pressure on Anthropic has dropped a layer, which both companies will internalize into their next-generation product roadmaps.

4. OpenAI Ships Bedrock Managed Agents on AWS the Day After the Microsoft Restructure

Within 24 hours of the Microsoft-OpenAI restructure announcement, AWS shipped a limited preview of OpenAI's frontier models, Codex, and a new Bedrock Managed Agents service on April 28 (AWS What's New, About Amazon, TechCrunch). The product surface mirrors what has been available on Azure for years: latest OpenAI models (presumably including GPT-5.5), Codex for code generation and execution, and the OpenAI agent harness wrapped in AWS's native enterprise governance — IAM, PrivateLink, CloudTrail, encryption at rest and in transit, and integration with existing compliance frameworks.

The timing speaks for itself. Either the AWS/OpenAI integration was developed in parallel with the Microsoft renegotiation and shipped the moment exclusivity dropped, or the partnership was structured tightly enough that an enterprise-grade Bedrock launch could be assembled in under a week — both readings point to OpenAI having actively engineered the post-exclusivity multi-cloud strategy rather than reacting to it. Andy Jassy's own description of the Microsoft-OpenAI restructure as a "very interesting announcement" the day before launch is consistent with a coordinated rollout.

The competitive consequence is that AWS now has both major proprietary frontier-AI vendors (Anthropic via Bedrock and OpenAI via Bedrock) running on its native enterprise stack. Bedrock Managed Agents is specifically optimized around OpenAI's reasoning models and agent harness, with each agent operating with its own identity and full CloudTrail logging — features that GA-level Anthropic-on-Bedrock offerings already provide for Claude. The result is that AWS's enterprise pitch is now "any frontier model with the same governance," while Azure's pitch is reduced to "Microsoft's preferred vendor relationship with OpenAI." For enterprises that have spent the last two years building Bedrock-anchored agent pipelines on Anthropic, the option to plug in OpenAI behind the same APIs is the largest single integration milestone of 2026 to date.

5. Google Commits Up to $40 Billion to Anthropic; Capital Arms Race Reaches $65 Billion in a Week

Google announced on April 24, 2026 a commitment of up to $40 billion in Anthropic — $10 billion immediate cash at the existing $350 billion valuation, with $30 billion contingent on undisclosed performance milestones (CNBC, Yahoo Finance/AFP, Intellectia analysis). Google Cloud also committed to deliver 5 gigawatts of compute capacity to Anthropic over five years beginning in 2027. This follows Amazon's commitment of up to $25 billion four days earlier, bringing Anthropic's confirmed funding commitments from the two largest cloud providers to a combined $65 billion plus multi-year compute infrastructure (Remio analysis).

The structural detail that matters is the 15% ownership cap. Google is structuring the equity so that its stake stays at or below 15% with no board seat, and Amazon's parallel structure is similar. Both hyperscalers are buying compute commitments and equity exposure without meaningful corporate-governance influence, leaving Anthropic's day-to-day strategic direction with Dario Amodei and the Anthropic board. The 5GW compute commitment is structurally more important than the cash: it represents the single largest discretionary AI compute commitment ever made to a single counterparty, and it begins flowing in 2027 at the same time Anthropic's existing AWS Trainium and Google TPU footprints are scheduled to step up significantly.

The financial context is Anthropic's growth: annualized revenue from $9 billion in December 2025 to $30 billion by April 2026, with enterprise customers spending over $1 million per year doubling from 500 to 1,000 in under two months. Both Google and Amazon are paying for proximity to a model that is, by Google's own internal assessment, currently winning the enterprise market that Google's Gemini also targets. The deeper strategic logic is that Google would rather own equity in the company eating its enterprise lunch than fail to participate, and Amazon's prior commitment forced Google's hand on timing. The next 12 months will reveal whether this is the high-water mark of hyperscaler-AI-lab capital intensity or merely the next plateau before a step-up driven by GPT-6 or Gemini-4 capabilities.

6. Apple Confirms Custom 1.2T-Parameter Gemini Powers Next-Generation Siri

Google Cloud CEO Thomas Kurian publicly confirmed at Cloud Next 2026 that Google is Apple's "preferred cloud provider" for the next generation of Apple Foundation Models, and that those models will be based on Gemini technology and will power the upcoming personalized Siri later in 2026 (MacRumors, Kingy AI analysis). Per WebProNews reporting cited in the Kingy coverage, the custom model is a 1.2 trillion-parameter Gemini variant built specifically for Apple's needs. Apple's announced two-phase launch puts limited Gemini-backed Siri features into iOS 26.4 (Spring 2026, already shipping) and the full conversational Siri into iOS 27 (September 2026 alongside the iPhone 18).

The strategic significance is the failure of Apple's prior "vertically integrated AI" thesis. Apple Intelligence as launched in 2024 was premised on a stack-up of on-device models plus a private-cloud fallback running Apple-trained models, with the explicit promise that Apple would not be reliant on third-party foundation models for core assistant capabilities. That promise has been quietly reversed: the on-device tier remains Apple's, but every interaction that requires frontier reasoning — multi-step task completion, cross-app context, conversational follow-up — is now routed to a Gemini variant Google trained, hosted on Google Cloud infrastructure. The reported architecture preserves user privacy via routing logic and on-device disambiguation, but the foundation-model layer is conceded.

The competitive read is that even the most aggressive vertical-integration story in consumer AI has chosen partnership over independent foundation-model development, after spending three years and significant capex trying. This is a meaningful data point for any organization currently considering "build versus partner" for frontier-model capabilities — Apple's resources, talent, and silicon advantage were among the strongest possible cases for "build," and the decision has gone the other way. For Google, the deal cements Gemini as a multi-platform frontier model and locks in distribution to roughly 1.5 billion active iOS devices, partially offsetting Anthropic's enterprise dominance and OpenAI's consumer dominance.

7. Meta and Microsoft Layoffs Tied Directly to AI Productivity Reallocation

Meta announced on April 23 a workforce reduction of roughly 10% — approximately 8,000 people — with another 6,000 open roles being eliminated, while Microsoft on the same day offered voluntary buyouts to roughly 7% of its US workforce, approximately 8,750 employees (CNBC, BBC, Axios). Layoffs.fyi tracking shows over 92,000 tech employees laid off in 2026 to date, on top of nearly 900,000 since 2020. Meta cited the cuts as part of a broader strategy to "enhance operational efficiency and counterbalance other investments" — those investments being the reported $135 billion AI capex run-rate for 2026, "roughly equivalent to what the company has invested in AI over the last three years combined."

The framing is unusually direct for industry communication. Meta's Janelle Gale memo explicitly tied workforce reduction to AI investment reallocation, and Microsoft offering voluntary buyouts for the first time in 51 years — a structural HR change, not a cyclical one — signals an internal model that anticipates structural rather than transient changes in labor demand. The industry pattern includes Snap (16% / 1,000 in March), Salesforce (4,000 customer-support roles in September 2025 with Marc Benioff explicitly saying "I need less heads"), Oracle (thousands in March citing AI investments), and Amazon's largest-ever cuts late last year.

The interesting structural question is which functions are absorbing the cuts. The available memos suggest middle management and operational/support layers are the primary targets, with technical roles in AI training, infrastructure, and product engineering protected and actively expanding. This matches the predictable shape of AI-driven labor substitution: roles that involve coordinating the work of others or executing routine knowledge tasks are first to go, while roles that produce inputs to AI systems (training data, evaluation, fine-tuning, infrastructure) are growing. The macro takeaway is that the "AI augments rather than replaces" framing of 2023-2024 has now visibly given way to direct substitution at scale at the largest tech employers, and the second-order effect on tech-adjacent labor markets — particularly contractor and BPO sectors that supplied many of the now-eliminated functions — will likely run through 2026 and 2027.

8. Qwen3.6-27B Dense Beats Alibaba's Own 397B MoE on Coding, Runs at 25 tok/s on a 24GB GPU

Alibaba's Qwen team released Qwen3.6-27B on April 22, 2026, the first dense open-weight model in the Qwen3.6 generation, under Apache 2.0 (NYU Shanghai RITS analysis, Hugging Face GGUF release). The headline result is that the dense 27B model outperforms Alibaba's own previous-generation Qwen3.5-397B-A17B Mixture-of-Experts flagship on multiple coding benchmarks — SWE-bench Verified at 77.2% vs 76.2%, with similar gaps on other agentic-coding evaluations — despite having approximately 14x fewer total parameters. The Q4_K_M GGUF quantization compresses BF16 weights from 55.6 GB to 16.8 GB, fitting on a single 24 GB consumer GPU, and Simon Willison's independent testing reports approximately 25 tokens/second generation via llama-server.

The architectural detail that should interest practitioners working on transformer alternatives is the model's hybrid: a Gated DeltaNet / Gated Attention architecture across 64 layers, with hidden size 5120 and FFN intermediate 17,408. This is the same Gated DeltaNet linear-attention pattern that Qwen3-Next pioneered at the 80B scale, now scaled down to a dense 27B that explicitly outperforms a much larger MoE on coding workloads. The mechanism is a 3:1 alternation between Gated DeltaNet (linear, sub-quadratic) and full attention layers, with the linear layers handling most of the long-range information transport and the full-attention layers providing the precise position-sensitive operations. The SWE-bench result is the strongest evidence to date that the linear-attention-mostly pattern transfers cleanly to demanding agentic coding tasks at small dense scale.

The deployment implication is that the gap between "frontier capability" and "running on a developer's workstation" has compressed materially. A model that matches Claude 4.5 Opus on Terminal-Bench-class tasks, sits within 3-4 points of frontier reasoning benchmarks, and runs on a single RTX 5090 at usable inference speed is no longer a research preview — it is a production option for any team with hardware access and a willingness to self-host. Combined with DeepSeek V4-Flash at the lower end of the Pro tier and GLM-5.1 at the higher end, the open-weight coding-model ecosystem now spans from consumer-GPU to multi-GPU MoE serving with frontier-comparable performance throughout.

9. Anthropic Investigates Possible Mythos Cybersecurity Model Breach

Anthropic is investigating a report of unauthorized access to its Claude Mythos preview model via a third-party vendor environment, the company confirmed on April 22, 2026 (BBC News, CBS News, Fortune). Bloomberg's prior reporting cited "a small group of unauthorized users" who accessed the model via a private Discord chat on the day Mythos was publicly announced. Mythos is the cybersecurity vulnerability detection model that Anthropic deliberately gated behind Project Glasswing, releasing access to roughly 50 partner organizations including Amazon, Apple, Cisco, JPMorgan Chase, and NVIDIA, on the explicit theory that the model's offensive capabilities — vulnerability discovery, automated exploitation reasoning — are too dangerous to release publicly.

Anthropic states it has not detected breaches outside the third-party vendor environment and has no evidence its own systems were compromised. The structural significance is that this is the first material-incident disclosure on a deliberately gated frontier model — a category that did not exist before Mythos. Project Glasswing was set up explicitly to thread the needle between "release a powerful defensive cybersecurity tool" and "do not give attackers access to a powerful offensive cybersecurity tool," and the architecture relies on the trustworthiness of a small number of vetted partners and their internal access controls. A breach via a third-party vendor environment is the predictable failure mode for that architecture, and the fact that it appears to have happened on the same day as the public announcement is uncomfortable.

The downstream implications are still developing. Anthropic has not disclosed which vendor environment was involved, what the unauthorized users accessed, or whether weights were exfiltrated. If weights were exfiltrated, the gating model collapses immediately, since open-weight Mythos is functionally indistinguishable from publicly released Mythos. If only API access was obtained, the harm surface is smaller but the precedent is established: gated-frontier-model access controls have a meaningfully higher attack surface than the model providers had been publicly modeling. For the broader frontier-model release calculus — which has been quietly shifting toward more "tiered access" and "verified deployment" patterns — the Mythos incident is the first stress test, and it has not gone well.

10. Two AI-in-Science Papers: New Physics from Dusty Plasma and the Emergence of Learning Mechanics

Two papers landed this week that bracket the scientific role of AI from opposite ends. The first is an Emory University PNAS paper on April 23 in which a physics-structured neural network discovered new non-reciprocal force laws in a dusty plasma — a "fourth state of matter" where ionized gas suspends charged dust particles — at greater than 99% accuracy (ScienceDaily report). The team, led by Justin Burton and Ilya Nemenman, used tomographic 3D imaging of particle trajectories combined with a structurally constrained neural architecture that decomposes motion into drag, environmental forces, and inter-particle forces. The model overturned two long-standing theoretical assumptions: that particle size does not affect the spatial decay of inter-particle forces, and that the standard exponential-decay form holds independent of particle properties. Both turned out to be wrong, and the AI revealed the corrections from limited experimental data — important because the small-data regime is the actual setting for most physics discovery.

The methodological contribution is that the network is "not a black box" — the architectural constraints are imposed by the physics, the latent variables are physically interpretable, and the discovered laws can be read off the trained weights and tested by independent experiment. This is the operational definition of "AI uncovers physical law" rather than "AI fits data," and the design is portable: Nemenman is teaching it at the Konstanz School of Collective Behavior on systems ranging from bird flocks to crowd motion. The constraint that makes the approach work — small data, structurally informed network, physically interpretable latents — is also the constraint that limits its generality, but the demonstration is significant for being the first end-to-end case of AI-driven physical-law discovery in a genuinely complex many-body system.

The second paper, posted to arXiv on April 26, is "There Will Be a Scientific Theory of Deep Learning" by Jamie Simon and a multi-institution team from UC Berkeley, Harvard, and NYU (arXiv 2604.21691). The paper argues that the field is in mid-transition from "alchemy" to a rigorous predictive science the authors call "learning mechanics," organized around five pillars: solvable idealized settings (deep linear networks), tractable limits (infinite-width regimes), simple empirical laws (neural scaling laws and the edge of stability), hyperparameter disentanglement, and universal behaviors. The argument is that these pillars are no longer disconnected research programs but are converging on a coherent body of theory analogous to early physics — predictive, mathematical, and capable of guiding experimental design. For practitioners, the practical implication is that "we don't really understand why deep learning works" is no longer a safe intellectual default, and the gap between "what we can train" and "what we can predict in advance" is narrowing in a way that will eventually feed back into training-recipe design and architecture search.

Cross-Cutting Themes

The week's signal is that 2024-2025 industry assumptions are weaker than they were seven days ago. The Microsoft-OpenAI exclusivity that anchored the cloud-AI commercial model is gone. The Apple "vertical integration" thesis that prevented Apple from depending on third-party foundation models is gone. The "frontier capability requires proprietary closed weights" framing is weaker after DeepSeek V4-Pro and Qwen3.6-27B. The "transformer is the architecture" framing is weaker after Qwen3.6-27B's Gated DeltaNet hybrid demonstrably outperforms a 14x-larger pure-attention MoE on coding. The "AGI clause is the central commercial trigger" framing is gone, replaced by bounded contractual structures with explicit caps and expirations.

What replaces these assumptions is a fully-meshed graph: every major lab is now reachable from every major cloud, hyperscaler equity stakes are capped at non-controlling positions, frontier capability lives in both proprietary and open-weight tiers, and the architecture frontier is meaningfully diverse. The capital intensity is still rising — $65B into Anthropic in a week, $135B in Meta capex, $5T in Nvidia market cap — but the structural diversity of where that capital is flowing has increased. For practitioners building on any of these technologies, the practical takeaway is that lock-in to any single vendor or architectural path is now a much weaker default than it was a year ago, and the rational position is to build agent and inference layers that can swap models across providers and parameter counts without rewriting the surrounding system.