arXiv:2605.16302v1 Announce Type: new Abstract: Reinforcement learning for multi-step reasoning with large language models (LLMs) often relies on sparse terminal rewards, leading to poor credit assignment conditions where the final feedback is evenly propagated across all intermediate decisions. This results in high gradient variance, unstable training, and numerous ineffective updates, ultimately causing the model to fail and preventing sustained improvement. We introduce a counterfactual comparison-based credit assignment framework, which samples multiple reasoning trajectories under the same input. By treating their differences as an implicit approximation of alternative decisions, we construct an implicit process-level advantage estimator that transforms sparse terminal rewards into step-sensitive learning signals. Based on this, we propose Implicit Behavior Policy Optimization (IBPO), which significantly improves training stability and performance upper bounds on mathematical and code reasoning benchmarks, pointing to a promising direction for unlocking the performance potential of LLMs.

arXiv:2605.16318v1 Announce Type: new Abstract: Building and maintaining state to learn policies and value functions is critical for deploying reinforcement learning (RL) agents in the real world. Recurrent neural networks (RNNs) have become a key point of interest for the state-building problem, and several large-scale reinforcement learning agents incorporate recurrent networks. While RNNs have become a mainstay in many RL applications, many key design choices and implementation details responsible for performance improvements are often not reported. In this work, we discuss one axis on which RNN architectures can be (and have been) modified for use in RL. Specifically, we look at how action information can be incorporated into the state update function of a recurrent cell. We discuss several choices in using action information and empirically evaluate the resulting architectures on a set of illustrative domains. Finally, we discuss future work in developing recurrent cells and discuss challenges specific to the RL setting.

arXiv:2605.16311v1 Announce Type: new Abstract: Distributed training of large neural networks is bottlenecked by full-precision gradient communication and by coordinatewise optimizers that ignore the matrix structure of weight tensors. We propose Sign-Muon, a 1-bit, matrix-aware optimizer that combines majority-vote sign aggregation from signSGD with the polar-step framework of Muon. Each worker forms a Muon-style direction by taking the polar factor of its momentum via a Newton--Schulz iteration, transmits only the entrywise signs, and aggregates by majority vote; an optional local polar step further enforces orthogonality at no extra communication cost. Under spectral-norm smoothness and bounded-variance stochastic gradients, the spectral-norm normalized sign step yields an $\mathcal{O}(1/\sqrt{T})$ nonconvex rate for an $\ell_1$-based stationarity measure. With unimodal symmetric noise, majority vote across $M$ workers cuts the stochastic term by $1/\sqrt{M}$, matching signSGD. In the $\alpha$-$\beta$ model, distributed Sign-Muon needs only one integer sum-allreduce per iteration; all orthogonalization is local, giving a $32\times$ bandwidth reduction over float32 ($4\times$ for int8). Across 330 CIFAR-10/ResNet-50 configurations Sign-Muon attains the best validation accuracy (92.15\%); its 4-GPU majority-vote variant reaches 92.02\% with 37\% less training time at matched effective batch. On nanoGPT, Sign-Muon achieves lower perplexity and better anytime performance than other sign-based baselines, with favorable weak-scaling up to 16 GPUs.

arXiv:2605.16650v1 Announce Type: new Abstract: Evaluating multi-turn dialogue systems remains challenging because response quality depends not only on the current prompt, but also on previously established entities, claims, and conversational commitments. Existing automatic evaluators, including LLM-as-a-judge frameworks and embedding-based metrics, largely rely on flat or turn-isolated representations, making them less effective at detecting long-range issues such as contradiction, topic drift, and entity inconsistency. To address this, we propose SKG-Eval, a quasi-deterministic and interpretable framework that models dialogue as an evolving Semantic Knowledge Graph (SKG) of entities, relations, and commitments across turns. The framework incrementally updates the graph through structured triple extraction and computes three complementary signals: (i) local relevance, measuring alignment with the current prompt and optional reference; (ii) historical consistency, evaluating how newly introduced information connects to prior conversational context using graph-based and embedding-driven signals; and (iii) logical coherence, assessed by a geometric contradiction engine that detects cross-turn conflicts without relying on NLI models or LLM judges. These signals are adaptively fused and aggregated into a length-invariant session score via recency-weighted trend analysis. Across multiple benchmarks, SKG-Eval achieves higher correlation with human judgments and substantially improves detection of long-range inconsistencies in extended conversations. In addition, the framework produces explicit contradiction certificates and deterministic scores for fixed inputs, enabling reproducible and auditable evaluation. Overall, our results suggest that structured externalized state tracking through semantic knowledge graphs provides a scalable alternative to implicit reasoning in LLM-based dialogue evaluators.

arXiv:2605.16262v1 Announce Type: new Abstract: Variational inequalities play a key role in machine learning research, such as generative adversarial networks, reinforcement learning, adversarial training, and generative models. This paper is devoted to the constrained variational inequality problems with functional constraints (inequality-type constraints). We propose some mirror descent-type algorithms that switch between productive and non-productive steps depending on the values of the functional constraints at iterations, with many different step size rules and stopping criteria. We analyze the proposed algorithms and prove their optimal convergence rate to achieve a solution with desired accuracy, for problems with bounded and monotone operators and Lipschitz convex functional constraints. In addition, we propose a modification of the proposed algorithms by considering each functional constraint in the calculation when we have a productive step, as well as the first constraint that violates the feasibility. This modification can save the running time of algorithms when we have many functional constraints. In addition, we provide an analysis of the proposed algorithms for $\delta$-monotone operators, allowing us to apply the proposed algorithms, as a special case, to constrained minimization problems when we do not have access to the exact information about the subgradient of the objective function. Numerical experiments that illustrate the work and performance of the proposed algorithms are also given.

arXiv:2605.16317v1 Announce Type: new Abstract: Accurate, unified models for event cameras (ECs) remain elusive, hampering calibration and algorithm design. We develop a foundational probabilistic model for EC event detection, grounded in photon statistics, that unifies the description of static scene noise events and step response curves (S-curves) within a single analytical framework. Three formulations of the probability distributions are derived, spanning all intensity regimes: exact Poisson, saddle-point, and Gaussian. The model reveals the underlying connection between these otherwise disparate EC behaviors and clarifies the interpretation of S-curves, which we show is more nuanced than selecting a fixed probability threshold. Based on this model, we propose Noise2Params, a method for determining camera-specific values of the log-contrast threshold $B$, the lux-to-photon conversion factor $\alpha$, and the leakage term $\theta$ (found to be intensity dependent), via error minimization against observed noise-event distributions. Noise2Params requires only recordings of static, uniform scenes, offering an experimentally accessible alternative to approaches that demand specialized dynamic light sources. We further support the validity the model by training convolutional neural networks (CNNs) on synthetic noise images generated from our distributions and evaluating their ability to reconstruct static scenes from experimental data. We further demonstrate the utility of our model by showing that CNNs incorporating synthetic data outperform those trained solely on experimental data. Our framework provides a quantitative foundation for EC calibration, noise-aware algorithm design, and applications in photon-limited regimes.

arXiv:2605.16679v1 Announce Type: new Abstract: End-to-end automation of realistic healthcare operations stresses three capabilities underrepresented in current benchmarks: policy density, decisions must be grounded in a large library of medical, insurance, and operational rules; Multi-role composition: a single task requires the agent to play multiple roles with handoffs; and multilateral interaction: intermediate workflow steps are multi-turn dialogs, such as peer-to-peer review and patient outreach. We introduce $\chi$-Bench, a benchmark of long-horizon healthcare workflows across three domains: provider prior authorization, payer utilization management, and care management. Each task hands the agent a clinical case in a high-fidelity simulator of 20 healthcare apps exposed via 87 MCP tools, which it must drive to a terminal status through tool calls and writing the role's artifacts, guided by a 1,290+ document managed-care operations handbook skill. Across 30 agent harness/models configurations, the best agent resolves only 28.0% of tasks, no agent clears 20% on strict pass^3, and executing all tasks in a single session slumps the performance to 3.8%. These results raise the hypothesis that similar gaps are likely to surface in other policy-dense, role-composed, irreversible enterprise domains.

Benedict Evans: An overview of macro tech trends: capex explosion, unprecedented surge in chip demand, supply chain bottlenecks, model commoditization, AI automation, and more  —  Twice a year, I produce a big presentation exploring macro and strategic trends in the tech industry.  —  New in May 2026, ‘AI eats the world’.

At Deutsche Börse Group, our StatistiX platform provides approximately 95% of all...

Bloomberg: Sources: Analog Devices, one of the largest analog chip makers, is in talks to buy Empower, which makes chips used to regulate voltage, for $1.5B in cash  —  Analog Devices Inc. is in advanced talks to acquire closely held Empower Semiconductor for $1.5 billion in cash, according to people familiar with the matter.

Although visitors to an event like TechEx North America will always want to see the cutting edge front and centre stage, the nuance and detail brought to the show by the speakers and exhibitors mean that it’s sometimes the smaller considerations that need to play big – at least, in the minds of enterprise decision-makers. […] The post AI is a matter of power, infrastructure and security: TechEx North America appeared first on AI News.

<p>I put together these annotated slides from my five minute lightning talk at PyCon US 2026, using the <a href="https://tools.simonwillison.net/annotated-presentations">latest iteration</a> of my <a href="https://simonwillison.net/2023/Aug/6/annotated-presentations/">annotated presentation tool</a>.</p> <div class="slide" id="5-minutes-llms.001.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.001.jpeg" alt="The last six months in LLMs in five minutes Simon Willison - simonwillison.net PyCon US 2026 Lightning Talk " style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.001.jpeg">#</a> <p>I presented this lightning talk at PyCon US 2026, attempting to summarize the last six months of developments in LLMs in five minutes.</p> </div> </div> <div class="slide" id="5-minutes-llms.002.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.002.jpeg" alt="The November inflection point " style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.002.jpeg">#</a> <p>Six months is a pretty convenient time period to cover, because it captures what I've been calling the <a href="https://simonwillison.net/tags/november-2025-inflection/">November 2025 inflection point</a>. November was a critical month in LLMs, especially for coding.</p> </div> </div> <div class="slide" id="5-minutes-llms.003.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.003.jpeg" alt="The “best” model changed hands 5 times between Anthropic, OpenAl and Google " style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.003.jpeg">#</a> <p>For one thing, the supposedly "best" model (depending mostly on vibes) changed hands five times between the three big providers.</p> </div> </div> <div class="slide" id="5-minutes-llms.004.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.004.jpeg" alt="Generate an SVG of a pelican riding a bicycle " style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.004.jpeg">#</a> <p>As always, I'm using my <a href="https://simonwillison.net/tags/pelican-riding-a-bicycle/">Generate an SVG of a pelican riding a bicycle</a> test to help illustrate the differences between the models.</p> <p>Why this test? Because pelicans are hard to draw, bicycles are hard to draw, pelicans <em>can't ride bicycles</em>... and there's zero chance any AI lab would train a model for such a ridiculous task.</p> </div> </div> <div class="slide" id="5-minutes-llms.005.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.005.jpeg" alt="Five pelicans, one for each of the following models. Varying qualities!" style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.005.jpeg">#</a> <p>At the start of November the widely acknowledged "best" model was Claude Sonnet 4.5, released on <a href="https://simonwillison.net/2025/Sep/29/claude-sonnet-4-5/">29th September</a>. It drew me this pelican.</p> <p>In November it was overtaken by <a href="https://simonwillison.net/2025/Nov/13/gpt-51/">GPT-5.1</a>, then <a href="https://simonwillison.net/2025/Nov/18/gemini-3/">Gemini 3</a>, then <a href="https://simonwillison.net/2025/Nov/19/gpt-51-codex-max/">GPT-5.1 Codex Max</a>, and then Anthropic took the crown back again with <a href="https://simonwillison.net/2025/Nov/24/claude-opus/">Claude Opus 4.5</a>.</p> <p>I think Gemini 3 drew the best pelican out of this lot, but pelicans aren't everything. Most practitioners will agree that Opus 4.5 held the crown for the next couple of months.</p> </div> </div> <div class="slide" id="5-minutes-llms.006.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.006.jpeg" alt="The coding agents got good " style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.006.jpeg">#</a> <p>It took a little while for this to become clear, but the real news from November was that the coding agents got <em>good</em>.</p> <p>OpenAI and Anthropic had spent most of 2025 running <a href="https://simonwillison.net/2025/Dec/19/andrej-karpathy/">Reinforcement Learning from Verifiable Rewards</a> to increase the quality of code written by their models, especially when paired up with their Codex and Claude Code agent harnesses.</p> <p>In November the results of this work became apparent. Coding agents went from often-work to mostly-work, crossing a quality barrier where you could use them as a daily-driver to get real work done, without needing to spend most of your time fixing their stupid mistakes.</p> </div> </div> <div class="slide" id="5-minutes-llms.007.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.007.jpeg" alt="Screenshot of &quot;Initial commit&quot; on GitHub to steipete/Warelay, commit f6dd362, steipete authored on Nov 24, 2025 It&#39;s a copy of the MIT license" style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.007.jpeg">#</a> <p>Also in November, this happened - the first commit to an obscure (back then) repo called "Warelay" by some guy called Pete.</p> </div> </div> <div class="slide" id="5-minutes-llms.008.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.008.jpeg" alt="December/January (A little bit of LLM psychosis) " style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.008.jpeg">#</a> <p>Over the holiday period, from December to January, a whole lot of us took advantage of the break to have a poke at these new models and coding agents and see what they could do.</p> <p>They could do a lot! Some of us got a little bit over-excited. I had my own short-lived bout of a form of LLM psychosis as I started spinning up wildly ambitious projects to see how far I could push them.</p> </div> </div> <div class="slide" id="5-minutes-llms.009.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.009.jpeg" alt="micro-javascript playground Execute JavaScript code in a sandboxed micro-javascript environment powered by Pyodide var numbers = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]; var doubled = numbers.map(n =&gt; n * 2); console.log(&#39;Doubled: &quot;&#39;, doubled); var evens = numbers.filter(n =&gt; n % 2 === 0); console.log(&#39;Evens: &#39;, evens); var sum = numbers.reduce((a, b) =&gt; a + b, @); console.log(&#39;Sum:&quot;, sum); Output 27 Doubled: [2, 4, 6, 8, 10, 12, 14, 16, 18, 20] Evens: [2, 4, 6, 8, 10] Sum: 55 Execution time: 8.00ms About: micro-javascript is a pure Python JavaScript interpreter with configurable memory and time limits. This playground runs entirely in your browser using Pyodide (Python compiled to WebAssembly). View on GitHub" style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.009.jpeg">#</a> <p>One of my projects was a vibe-coded implementation of JavaScript in Python - a loose port of <a href="https://github.com/bellard/mquickjs">MicroQuickJS</a> - which I called <a href="https://github.com/simonw/micro-javascript">micro-javascript</a>. You can try it out in your browser in <a href="https://simonw.github.io/micro-javascript/playground.html">this playground</a>.</p> </div> </div> <div class="slide" id="5-minutes-llms.010.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.010.jpeg" alt="JavaScript running in Python running in Pyodide running in WebAssembly running in JavaScript" style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.010.jpeg">#</a> <p>That playground demo shows JavaScript code run using my micro-javascript library, in Python, running inside Pyodide, running in WebAssembly, running in JavaScript, running in a browser!</p> <p>It's pretty cool! But did anyone out there <em>need</em> a buggy, slow, insecure half-baked implementation of JavaScript in Python?</p> <p>They did not. I have quite a few other projects from that holiday period that I have since quietly retired!</p> </div> </div> <div class="slide" id="5-minutes-llms.011.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.011.jpeg" alt="February 2026 " style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.011.jpeg">#</a> <p>On to February. Remember that Warelay project that had its first commit at the end of November?</p> </div> </div> <div class="slide" id="5-minutes-llms.012.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.012.jpeg" alt="Warelay → CLAWDIS → CLAWDBOT → Clawdbot → Moltbot →🦞 OpenClaw" style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.012.jpeg">#</a> <p>In December and January it had gone through <a href="https://simonwillison.net/2026/May/16/openclaw-names/">quite a few name changes</a>... and by February it was taking the world by storm under its final name, <a href="https://openclaw.ai/">OpenClaw</a>.</p> <p>The amount of attention it got is pretty astonishing for a project that was less than three months old.</p> </div> </div> <div class="slide" id="5-minutes-llms.013.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.013.jpeg" alt="Generic term: Claw " style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.013.jpeg">#</a> <p>OpenClaw is a "personal AI assistant", and we actually got a generic term for these, based on NanoClaw and ZeroClaw and suchlike... they're called <strong>Claws</strong>.</p> </div> </div> <div class="slide" id="5-minutes-llms.014.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.014.jpeg" alt="An aquarium for your Claw " style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.014.jpeg">#</a> <p>Mac Minis started to sell out around Silicon Valley, because people were buying them to run their Claws.</p> <p><a href="https://www.dbreunig.com/">Drew Breunig</a> joked to me that this is because they're the new digital pets, and a Mac Mini is the perfect aquarium for your Claw.</p> </div> </div> <div class="slide" id="5-minutes-llms.015.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.015.jpeg" alt="Alfred Molina&#39;s Doc Ock in Spider-Man 2, tearing apart a New York subway train with his four claws." style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.015.jpeg">#</a> <p>My favourite metaphor for Claws is Alfred Molina's Doc Ock in the 2004 movie Spider-Man 2. His claws were powered by AI, and were perfectly safe provided nothing damaged his inhibitor chip... after which they turned evil and took over.</p> </div> </div> <div class="slide" id="5-minutes-llms.016.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.016.jpeg" alt="Gemini 3.1 Pro A really good illustration of a pelican riding a bicycle. " style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.016.jpeg">#</a> <p>Also in February: Gemini 3.1 Pro came out, and drew me a <em>really good pelican riding a bicycle</em>. Look at this! It's even got a fish in its basket.</p> </div> </div> <div class="slide" id="5-minutes-llms.017.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.017.jpeg" alt="Gemini 3 Pro pelican contrasted with Gemini 3.1 Pro, as animated SVGs" style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.017.jpeg">#</a> <p>And then Google's Jeff Dean <a href="https://simonwillison.net/2026/Feb/19/gemini-31-pro/#jeff-dean">tweeted this video</a> of an animated pelican riding a bicycle, plus a frog on a penny-farthing and a giraffe driving a tiny car and an ostrich on roller skates and a turtle kickflipping a skateboard and a dachshund driving a stretch limousine.</p> <p>So maybe the AI labs have been paying attention after all!</p> </div> </div> <div class="slide" id="5-minutes-llms.018.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.018.jpeg" alt="April 2026 " style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.018.jpeg">#</a> <p>A lot of stuff happened just in the past month.</p> </div> </div> <div class="slide" id="5-minutes-llms.019.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.019.jpeg" alt="Gemma 4 26B-A4B (17.99GB) A pretty decent pelican riding a bicycle, though the bike is a bit mis-shapen." style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.019.jpeg">#</a> <p>Google released the <a href="https://simonwillison.net/2026/Apr/2/gemma-4/">Gemma 4</a> series of models, which are the most capable open weight models I've seen from a US company.</p> </div> </div> <div class="slide" id="5-minutes-llms.020.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.020.jpeg" alt="GLM-5.1 MIT, 754B parameter, 1.51TB! " style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.020.jpeg">#</a> <p>Also last month, Chinese AI lab GLM came out with <a href="https://simonwillison.net/2026/Apr/7/glm-51/">GLM-5.1</a> - an open weight 1.5TB monster! This is a very effective model... if you can afford the hardware to run it.</p> </div> </div> <div class="slide" id="5-minutes-llms.021.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.021.jpeg" alt="" style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.021.jpeg">#</a> <p>GLM-5.1 drew me this very competent pelican on a bicycle.</p> </div> </div> <div class="slide" id="5-minutes-llms.022.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.022.jpeg" alt="The bike is wonky, the pelican is floating." style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.022.jpeg">#</a> <p>... though when it <a href="https://gisthost.github.io/?73bb6808b18c2482f66e5f082c75f36e">tried to animate it</a> the bicycle bounced off into the top and the bicycle got warped.</p> </div> </div> <div class="slide" id="5-minutes-llms.023.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.023.jpeg" alt="Screenshot of Bluesky Charles ‪@charles.capps.me‬ I think you should pester it with another animal using another method of locomotion. Something tells me it was trained for this. I can&#39;t quite put my finger on it. /s NORTH VIRGINIA OPOSSUM ON AN E-SCOOTER!!" style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.023.jpeg">#</a> <p>Charles <a href="https://bsky.app/profile/charles.capps.me/post/3miwrn42mjc2t">on Bluesky</a> suggested I try it with a North Virginia Opossum on an E-scooter</p> </div> </div> <div class="slide" id="5-minutes-llms.024.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.024.jpeg" alt="NORTH VIRGINIA OPOSSUM CRUISING THE COMMONWEALTH SINCE DUSK And a really cool illustration of a possum." style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.024.jpeg">#</a> <p>And it did this! I've tried this on other models and they don't even come close. "Cruising the commonwealth since dusk" is perfect. It's <a href="https://static.simonwillison.net/static/2026/glm-possum-escooter.html">animated too</a>.</p> </div> </div> <div class="slide" id="5-minutes-llms.025.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.025.jpeg" alt="Qwen3.6-35B-A3B is a 20.9GB file that runs on my laptop It drew a better pelican on a bicycle than Opus 4.7, which messed up the bicycle frame." style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.025.jpeg">#</a> <p>The other neat Chinese open weight models in April came from Qwen. <a href="https://simonwillison.net/2026/Apr/16/qwen-beats-opus/">Qwen3.6-35B-A3B on my laptop drew me a better pelican than Claude Opus 4.7</a>. That's a 20.9GB open weights model that runs on my laptop!</p> <p>(I think this mainly demonstrates that the pelican on the bicycle has firmly exceeded its limits as a useful benchmark.)</p> </div> </div> <div class="slide" id="5-minutes-llms.026.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.026.jpeg" alt="Claude Sonnet 4.5 pelican for comparison." style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.026.jpeg">#</a> <p>Here's that Claude Sonnet 4.5 pelican from September for comparison. </p> </div> </div> <div class="slide" id="5-minutes-llms.027.jpeg"> <img loading="lazy" src="https://static.simonwillison.net/static/2026/5-minutes-llms/5-minutes-llms.027.jpeg" alt="The themes of the past 6 months: Coding agents got really good Local models wildly outperform expectations " style="max-width: 100%" /> <div><a style="float: right; text-decoration: none; border-bottom: none; padding-left: 1em;" href="https://simonwillison.net/2026/May/19/5-minute-llms/#5-minutes-llms.027.jpeg">#</a> <p>So those were the two main themes of the past six months. The coding agents got really good... and the laptop-available models, while a lot weaker than the frontier, have started wildly outperforming expectations.</p> </div> </div> <p>Tags: <a href="https://simonwillison.net/tags/lightning-talks">lightning-talks</a>, <a href="https://simonwillison.net/tags/pycon">pycon</a>, <a href="https://simonwillison.net/tags/speaking">speaking</a>, <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/local-llms">local-llms</a>, <a href="https://simonwillison.net/tags/llms">llms</a>, <a href="https://simonwillison.net/tags/annotated-talks">annotated-talks</a>, <a href="https://simonwillison.net/tags/pelican-riding-a-bicycle">pelican-riding-a-bicycle</a>, <a href="https://simonwillison.net/tags/coding-agents">coding-agents</a></p>

Discussion | Link

Jason Nelson / Decrypt: Vitalik Buterin says AI-assisted “formal verification” could help secure blockchain networks, smart contracts, and cryptographic systems against software flaws  —  Ethereum co-founder Vitalik Buterin said that mathematically verified software is becoming essential to protecting Ethereum …

On Monday, the jury in Musk v. Altman dealt Elon Musk a major blow—reaching a unanimous advisory verdict that he had sued OpenAI too late and, as a result, his claims are barred by the applicable statutes of limitations. US District Judge Yvonne Gonzalez Rogers immediately accepted it.  Musk announced on X that he will…

Discussion | Link

Researchers at Penn have created a hybrid light-matter particle that could dramatically speed up AI computing while using far less energy. The breakthrough may help replace some electronic computing processes with ultra-efficient light-based technology.

Wall Street Journal: Blackstone announces a JV with Google to create a US company that will offer customers access to Google's TPUs, and makes an initial equity commitment of $5B  —  Investment firm to put $5 billion toward venture using Google's chips  —  Alphabet's Google and Blackstone plan to create …

Riley Griffin / Bloomberg: A deep dive into Meta's Hyperion data center campus in Louisiana, expected to reach 5 GW of compute capacity; source: Meta plans to spend $200B+ on the project  —  Dustin Morris steers his propeller plane over farmland shrouded in a ghostly morning mist.  From 900 feet up, Richland Parish …