In this article
OpenAI says its models found up to 14% improvement in power, performance, and area, or PPA, on selected blocks of AMD’s flagship GPU IP in less than five months. Some proposed changes entered a design sent for manufacturing. The same program produced single-digit PPA gains on AMD’s next flagship CPU and cut standard-cell area by an estimated 13% in the vector unit of OpenAI’s Jalapeño inference chip. (OpenAI)
That is a production signal, not a whole-chip performance claim. OpenAI does not identify the blocks, publish the PPA weighting, or provide an independent audit. Its commercial position also matters: OpenAI and AMD have a multiyear compute agreement that includes warrants for up to 160 million AMD shares. (AMD) The public evidence supports a narrower conclusion. A model can now propose useful RTL changes inside an industrial verification loop, but the loop decides which changes count.
OpenAI’s open RISC-V evaluation makes that boundary visible. Across 9,881 scored runs, five model generations tried two tasks on the PicoRV32 and Ibex cores. One task optimized existing RTL. The other designed a core from written requirements. Candidates received zero if simulation failed or area exceeded 16 times the original core. Passing designs were scored using hidden workloads, instructions per clock, and a Yosys estimate of maximum clock frequency.
At a comparable budget of about 112,000 tokens per run, GPT-6.2 Astra led both cores. More revealingly, every model scored higher when starting from requirements. OpenAI says part of that advantage came from freedom to omit features in the original cores, including performance counters and additional instructions. A smaller design can win when the score does not charge it for missing behavior.
For a principal engineer, this changes the question from “Can the model optimize the design?” to “Can the release system distinguish optimization from specification loss?”
Make the contract executable before making the optimizer autonomous
The chip workflow suggests four boundaries that transfer to software agents.
- Separate invariants from objectives. Functional equivalence, required instructions, security properties, and interface behavior are hard constraints. Area, latency, cost, and throughput are objectives to optimize only after those constraints pass.
- Use a fidelity ladder. Simulation can reject obvious failures. Logic synthesis tests timing and area assumptions. Physical design checks routing, congestion, and power. Silicon closes the final gap. An agent should not promote evidence from one rung as proof of the next.
- Preserve a reviewable delta. Store the requirement version, proposed RTL, tool outputs, failed attempts, and the reason each accepted change improved the objective. A final score without the design trajectory is weak release evidence.
- Keep sign-off independent from search. The model that explores changes should not define the acceptance tests or waive failures. OpenAI’s Synopsys partnership is built around agents operating EDA tools, while established verification and engineer review remain the ground-truth boundary. (Synopsys)
The architectural unit is therefore not the model. It is the search process plus executable constraints, staged verification, and accountable sign-off.
Application radar: Meta uses agents to propagate proven changes
Meta published a production example of the same pattern in a different domain. Its ads-ranking organization moves proven techniques across a fleet of models with different architectures, data, and serving constraints. Meta says agents have driven execution since March 2026 and now handle more than 60% of technique propagation. Early results include three times more adaptations per engineer, a median fourfold reduction in training compute across selected technique-model combinations, and an estimated $45 million or more in 2026 compute savings. (Meta)
The important prerequisite came before the agents. Meta says standard templates and deterministic automation had already shipped more than 5,000 technique-model combinations while reducing per-model engineering time by 92%. Agents operate on top of that structure. They rank candidate transfers, adapt a technique, manage multiday experiments, diagnose failures, and reclaim compute from low-value runs. A shared record of prior runs supplies the evidence base.
Release remains gated. Candidate selection is measured with precision and recall. Adaptations first run on scaled-down models, then representative full-size models. They must improve offline predictive metrics without regressing training or inference cost, and they need a statistically significant online A/B win before shipping. Humans review key decisions and approve the final proposal.
The published figures are Meta’s own measurements, and the fourfold compute result applies to selected combinations rather than the full fleet. Even with that limitation, the case provides a strong design lesson: the agent scales a pre-existing evaluation and deployment system. It does not replace one.
Deadline radar: 2 December for older generative systems in the EU
A near-term control deadline turns output provenance into an operational requirement. Article 50(2) of the EU AI Act requires covered providers of systems that generate synthetic audio, image, video, or text to mark outputs in a machine-readable format that is detectable as artificially generated or manipulated. Systems placed on the market before 2 August 2026 received a transition period that ends on 2 December 2026. (European Commission)
This is not a blanket statement that every AI-assisted artifact needs the same label. Scope, exceptions, provider versus deployer duties, and editorial-control provisions differ. Teams should use the Commission’s guidance and legal review for their facts. The engineering work is still concrete:
- inventory production generators and their market-entry dates;
- record where marking is added, transformed, or stripped across the delivery chain;
- test the marker after common export, compression, and republishing paths;
- assign an owner for exceptions and evidence;
- retain versioned proof that a specific output path met the applicable rule.
The deadline is close enough that a policy document without an end-to-end conformance test is a delivery risk.
Watch next
- Whether OpenAI or AMD publishes block identities, PPA weighting, physical-design results, or silicon measurements for the changed AMD blocks.
- Whether Meta reports full-fleet denominators, failure rates, and online experiment outcomes beyond the selected technique-model combinations.
- Whether the Commission’s final implementation material changes the Article 50(2) scope or the evidence expected before 2 December.
Stay Sharp: mutate the evaluator, not only the implementation
Mutation testing evaluates a test suite by inserting deliberate faults into code. If the tests fail, the mutation is killed. If they still pass, the suite may not observe behavior it claims to protect. In 2021, Google’s mutation-testing program applied the technique incrementally during code review and filters unproductive mutations rather than chasing raw volume. (Google Testing Blog)
Apply the same idea to an agent or optimization contract. Before trusting the evaluator, create controlled defects that should be rejected:
- delete a required instruction or API field;
- weaken an authorization boundary;
- remove a consent disclosure;
- change a timeout or retry limit;
- preserve the final answer while corrupting an intermediate artifact.
Run the release suite against each mutant. A surviving mutation identifies a blind spot in the contract, the instrumentation, or the assertions. Track mutation results by risk class, not as one vanity percentage. Some mutations are equivalent or irrelevant, so review them instead of forcing a 100% score.
This is the practical response to the RISC-V result. A benchmark can execute every path and still fail to notice that a useful feature vanished. Mutation testing asks a harder question: if the optimizer exploits an omission, will the release system catch it?
The most capable optimizer will find empty space in its contract. Architecture earns trust by making that space observable before production does.