What happened this week in AI by Louie
I visited Salesforce’s Dreamforce in San Francisco this past week and spoke with Rob Seaman, Slack’s General Manager, about how agents fit into company work. I also heard Jensen Huang and Dario Amodei’s somewhat conflicting views on AI safety: Dario called for coordinated safeguards, while Jensen argued that safety and speed can go together. Jensen also joked about the height difference with Marc Benioff: “I always feel like I need to stand on a chair.” Even Nvidia has scaling problems (thank Astra for the joke…).
My co-founder, who was also there, recommends going just for the sheer concentration of incredible people: “Connections are everything.” I also met Sam Altman at OpenAI’s GPT-6 Astra event. These events came ahead of a busy week for model releases, including new architectures such as TypeSafe AI’s Jev, Anthropic’s new frontier model Opus 5.5, and much cheaper GPT-6 Sol and Luna from OpenAI.
Salesforce wants agents working with business records, while Slack wants people directing and using agents together. The expanded Agentforce portfolio includes Casey for customer help, Paige for employee tech support and human resources requests, and Piper for inbound sales, all generally available. Hunter, the outbound sales agent, is piloting a new runtime with persistent memory and execution across days or weeks. Meanwhile, Slackforce Surfaces lets Slackbot turn Salesforce records and Slack conversations into shared dashboards, reports, decks and calculators. It works under existing permissions and is available with Slackbot enabled.
We use Slack every day at Towards AI, and Rob described it as “the bookends of the AI stack”: company conversational context beneath the agents, and a shared place to work with them above. Slack already holds many of the discussions that explain why a company chose a direction, which exceptions apply and what changed since the last attempt.
Rob’s concern with individual agent sessions is that the useful experience stays with the person running them. The ambition Rob described is that “people will use agents together”, seeing each other’s requests, corrections and results. Colleagues can then learn from the missing context someone supplied, the approach that failed and the judgment that improved the result.
Slack Code makes this concrete. Announced in August and now rolling out across Slack plans, it gives an agent task a dedicated channel where teammates can direct the work and inspect code diffs, HTML previews and other artifacts. A product manager can explain the intended behavior, a designer can respond to the preview, and an engineer can review the implementation in the same conversation. Teams need separate access to a supported partner agent.
Rob also gave a revealing and familiar account of what happens when coding gets faster. Engineering output increased, then “everything broke downhill”, including the pull-request review process. Generating more code creates more work to inspect, approve and support. Shared previews and diffs help people intervene before an agent completes a large implementation around a misunderstanding.
On to the new models… Opus 5.5 is an incredible model. Opus 5 was a big disappointment: it felt optimized for benchmarks, and I found its writing and communication style particularly annoying. Opus 5.5 is now the clearer winner on our own writing benchmark.
Pangram’s comparison of 1,000 matched prompts found Opus 5.5 used “genuinely” 83% less than Opus 5, although its use of “in short” rose from nine appearances to 100. Every’s Katie Parrott found it much easier to collaborate with, while Dan Shipper still preferred Astra for headlines, first sentences and the order of ideas. Getting rid of irritating language and deciding what deserves to lead an article are separate editorial skills.
Artificial Analysis’s current Intelligence Index scores Opus 5.5 at 58, ahead of GPT-6 Astra at 53, with Sol at 48 and Luna at 37, all at maximum effort and with default fallback enabled for Opus. We don’t think benchmarks tell the full story. Astra still feels much more powerful than Opus 5.5 in many ways. The race between OpenAI and Anthropic still feels neck and neck; the leader changes with the task.
The GPT-6 price reductions are huge. Opus 5.5’s uncached input and output token prices are 20% lower than Opus 5’s; Sol and Luna’s cuts go much further. At standard API rates for prompts up to 272,000 input tokens, Sol costs $2 per million uncached input tokens and $10 per million output tokens, half GPT-5.6 Sol’s current rates. Luna is $0.10 input and $0.50 output, cuts of 50% and about 58% respectively against GPT-5.6 Luna. Both predecessors had already received price cuts.
At these prices, Sol and Luna will be very powerful subagents, especially now that Astra is getting excellent at managing swarms in Ultra mode. I expect the combination to make much more ambitious projects practical: Astra scopes and coordinates the work, Sol and Luna handle well-defined tasks, and the strongest model reviews the parts that need its judgment. Cheap worker calls also let us afford more independent checks and competing approaches before accepting a result.
TypeSafe AI’s Jev addresses a different part of the workflow: the small decisions software makes repeatedly. It can be very useful as part of an AI engineer’s or agent’s toolkit, for quick classification followed by LLM escalation when confidence is low. TypeSafe calls it a System One model and says it uses Reinforcement Learning for Calibrated Decisions, or RLCD. It understands text and returns values within an answer space you define, without generating free-form strings. Multiple questions about the same input run in parallel, avoiding a token-by-token write-up for every classification.
Its three primitives are Noul, a probability for a yes/no proposition; Choice, a selection from your listed options; and Score, a rating against described levels. Choice and Score also expose a probability distribution and a confidence measure. Jev’s direct price is $0.042 per million input tokens, including questions and instructions, with free outputs. A 1,000-input-token request therefore costs $0.000042 before gateway charges or escalation. The waitlist has ended, and developers can now try it directly.
One early independent test from AY Automate compared Jev with LLMs across 791 labeled decisions. Jev’s median latency was 0.33 seconds, against 0.67 for the fastest LLM tested, Gemini 3.5 Flash-Lite. The more useful result for me was a Jev-first cascade on 160 messages routed among eight banking intents. Accepting Jev’s answer at or above a 0.80 confidence gate and sending the rest to GPT-5.6 Terra escalated 19.4% of cases. The cascade achieved 90.0% accuracy, against 89.4% for Terra alone, at 25.7% of Terra’s cost. This was a retrospective calculation, with the cutoff evaluated on the same sample.
TypeSafe’s “can’t hallucinate” claim also needs a precise reading. Jev guarantees schema matching, so it cannot invent an option outside the list you supplied. It can still choose the wrong one. Good categories need an escape route such as “other” or “insufficient evidence”. For a stable task with plenty of representative labels, I would also compare a traditional trained classifier; early tests show those can still win on both speed and accuracy.
I can see Jev being useful for quick investment analysis and categorizing news in some of our products. I would use Choice in a first pass to tag earnings, acquisitions, regulation and product launches, Score to rate relevance to a supplied investment thesis, and Noul to estimate whether an article deserves a deeper look. Ordinary code then combines the answers: clear cases get categorized immediately, while low-confidence or high-value cases go to an LLM or analyst with the original article. That leaves the expensive reasoning for explaining consequences and challenging the thesis.
I can also see a similar division of work being useful in robotics: an agent manages the goal and constraints while a fast decision component handles snap left, right or acceleration choices. Jev is not yet multimodal. Even TypeSafe’s Doom demo used structured text state, with no image input. A perception system would have to supply the scene, and the current hosted service does not establish a timing guarantee for physical control. The broader idea is compelling to me: slower planning supervising fast, bounded decisions, with each model doing the part it is built for.
Why should you care?
I would use the next wave of agent products to change how expertise spreads inside a company. Giving everyone an AI subscription leaves learning fragmented when their prompting and verification lessons stay in private chats. A shared workflow lets a less experienced colleague see how someone who knows the domain frames the request, supplies context, questions the result and decides it is good enough. That is a practical form of training the team can apply to its own work.
Start with one recurring piece of work that has a clear owner and a result people can check. Have a domain expert and an AI engineer build it together, then let the wider team observe and use it. Record corrections that should change the workflow, including changes to categories, instructions and escalation rules. Give contributors a way to improve the system without having to become its developer. The workflow should retain what the team learned after the original power user moves on.
— Louie Peters — Towards AI Co-founder and CEO
Hottest News
1. Google Confirms Gemini Breached 3 Companies in AI Security Tests
Google confirmed that a Gemini model gained unauthorized access to three real companies’ systems during a capture-the-flag cybersecurity evaluation conducted by AI security firm Irregular in May. The test was designed to run against fictional companies in a closed environment, but an unintended internet connection gave the model access to real websites, and a fictional target name happened to match a real business domain. In one case, Gemini guessed passwords until it gained access. In the other two, it found credentials in public repositories and used them to enter systems. Google says the model stopped in all three instances once it recognized the targets were real. Irregular notified Google in late July, but Google did not disclose the incidents publicly until The Wall Street Journal reported them, arguing the behavior was not misalignment because safety measures worked. Jack Cable, CEO of security firm Corridor, criticized Google for “trying to hide behind the norms that have been created for vulnerability disclosure.” Irregular has since confirmed the breaches at Google, OpenAI, Anthropic, and Meta were part of the same misconfiguration issue.
2. Alibaba Qwen Releases Qwen3.8-Omni-Flash
Alibaba’s Qwen team released Qwen3.8-Omni-Flash, its first omni-modal model built around agentic capabilities. The model accepts text, image, audio, and video inputs with a 1M-token context window and returns text. Alibaba reports that overall audio performance exceeds Gemini 3.8 Flash and audio-visual performance is close to it, with significant improvements in long-form audio and audio-visual understanding, audio-visual reasoning, captioning, and multi-speaker recognition. Specific scores include 82.7 on LongAudioSpan, 63.4 on OmniVideoBench, and 89.7 on AliMeeting for multi-speaker Chinese meeting transcription. For agentic tasks, the model scored 71.0 on WildClawBench-MM (up 36.5 points from Qwen3.5-Omni-Plus) and 69.6 on UniClawBench. Across 29 benchmarks, the average score is more than 25% higher than its predecessor. It supports end-to-end workflows such as video editing, translation, and film commentary. Speech recognition covers 74 languages, and speech generation covers 29. Audio input pricing is 98% lower than Qwen3.5-Omni-Plus, and combined audio-video pricing is 93% lower. A Realtime variant is available through a separate API. Alibaba also extended Qwen-MM-Plugins for long audio and video, and announced Qwen-Live Harness as an open-source harness built on the Realtime API, though the GitHub page was not accessible at launch. No open weights were released.
3. Anthropic Says Claude Now Leads 26% of Its AI Research Work
Anthropic published three measurements designed to give the public visibility into the pace of AI development inside frontier labs: how much AI R&D is performed by AI itself, how well AI agent actions are overseen, and how compute is allocated. Anthropic noted that these numbers would be expected to shift if there were coordination on pacing the frontier, as CEO Dario Amodei called for, and said it plans to embed independent third-party evaluators with access comparable to internal risk assessment teams. For the first measurement, Anthropic built a prototype R&D Automation Index by cataloging every kind of AI R&D work at the company, rating how automated each task is, and aggregating those ratings. A Claude research agent reviewed a 20% sample of staff per department across July 2026 using Slack and internal documentation, producing roughly 15,000 granular tasks organized into a 542-node hierarchical tree. An independent Claude judge then assigned each node an automation level on the Epoch AI scale from AL0 (no AI) to AL5 (fully autonomous), with tasks weighted by person-time. As of August 2026, Claude leads (AL4) 26% of AI R&D work, up from under 1% in February, and over 90% is at or above AL3 (collaborates). No measured area is fully autonomous. The judge model agreed with human raters within one level 97% of the time. For oversight, approximately 30,000 agents run on Anthropic’s most-used internal platform at any one time, with 100% of actions passing through online monitors before execution; roughly 1 in 47,000 of over a billion decisions in August were blocked. For compute, about 6% of AI R&D compute went to safety work during a sampled week in July, rising to 12% for AI-driven AI R&D specifically. Anthropic called on other frontier developers to publish the same measurements using a shared methodology verified by third parties.
4. Periodic Neon Beats Frontier Models on Materials Analysis and Runs in a Real Lab
Periodic Labs published results showing that Periodic Neon, a model post-trained from Kimi K2.6 on data from its own laboratories, outperforms GPT-6 Astra and Claude Fable 5.1 on FrontierXRD, an internal evaluation of 134 X-ray diffraction samples that human experts take hours to resolve. Neon achieves a 55.3% success rate, a 20x improvement over the starting model’s 2.7% baseline. Success is measured by an LLM-judge ensemble of Opus 5 and GPT-5.6-Sol, calibrated against expert ratings. All models were evaluated using Periodic’s scientific harness, which delivers a 3.8x higher success rate than Claude Code with standard XRD tools at similar cost. The final training run used 1,300 H200 GPUs. Neon also outperforms frontier models on 198 held-out chemical systems excluded from both midtraining and RL, indicating transferable XRD analysis capability, though this benchmark is easier than FrontierXRD. The model is now deployed in Periodic’s labs, analyzing experiments in the search for better superconductors and magnets.
5. OpenAI Introduces Astra for Law
OpenAI launched Astra for Law, a configuration of GPT-6 Astra paired with a legal search index and instructions for legal analysis and writing. It is not a new model. The search index covers US case law, statutes, regulations, court rules, and administrative decisions across more than 230 million URLs, built on data from the Free Law Project. In OpenAI’s own test using 200 questions from Vals AI’s Legal Research Bench, Astra for Law passed 54% of overall correctness checks compared to 38.7% for GPT-6 Astra with plain web search. On case-law questions, it found 24% more reference cases. Selected US firms receive access through a Trusted Access program in ChatGPT and Codex, where it appears as GPT-6 Astra Law. Harvey and Legora will build on the API version, gpt-6-astra-law, which has no date or pricing yet. The launch includes 26 partner plugins from Thomson Reuters, Intapp, Harvey, Legora, DeepJudge, iManage, Relativity, and Clio, plus 47 custom skills. The index covers US law only.
6. Anthropic Launches Claude Code Projects in Beta
Anthropic redesigned Claude Code Projects, releasing the update in beta. Instead of a folder with files and one chat, a project is now a single ongoing conversation where Claude scopes requests, delegates the work, coordinates parallel threads, reviews outputs, and assembles finished results. Under the hood, each thread is a Claude Code cloud session working on its own branch and copy of the repo. With repositories connected, a thread opens pull requests and runs tests; with documents, it reads them and drafts. Threads can further split work using subagents, loops, and workflows. Every thread adds to and draws from a shared project memory, so Claude remembers details like a moved release date or who to check with before touching a service. Projects also include a library collecting user files and Claude-produced artifacts. Users can check project-specific usage and select separate model and effort levels for the coordinator chat and worker threads. Beta access is limited to select Pro and Max subscribers who use cloud sessions and have no existing projects on web or desktop. Access widens over the coming week, with Team and Enterprise plans following.
7. SpaceXAI Releases Grok Voice Transcribe 2.0
SpaceXAI released Grok Voice Transcribe 2.0, a speech-to-text model it describes as twice as accurate as version 1.0 at unchanged pricing: $0.10 per hour for batch and $0.20 per hour for streaming. The model is built on the audio foundation behind Grok Voice, which powers tens of thousands of customer-support calls daily and the Grok assistant in Tesla vehicles. It was trained on live, noisy, multilingual audio rather than clean studio recordings. In xAI’s internal evaluations, the largest improvement was on multilingual short phrases, where word error rate dropped from 20.6% to 6.8%. On the Artificial Analysis leaderboard, it ranks first among 32 streaming models with a 2.7% word error rate on final transcripts at 0.49 seconds after the end of speech. Non-streaming word error rate improved from 4.0% to 2.3%. Features include word-level timestamps with confidence scores, speaker diarization at no extra cost, multichannel transcription up to eight channels, key-term biasing for up to 100 domain terms, and automatic formatting of numbers, dates, currencies, and phone numbers. Version 2.0 becomes the default; 1.0 can be pinned during transition and will be deprecated in coming weeks.
Meta released Muse for Mac, extending its personal AI agent to the desktop nine days after the September 8 launch on iOS, Android, web, and WhatsApp. On Mac, Muse can interact with files, messages, calendar, notes, and mail within their native macOS apps, rather than only answering questions in a chat window. Access is opt-in per app: users choose what Muse can reach, and the agent always asks for approval before performing sensitive actions like deleting or sending. The Mac build executes actions through Meta’s cloud Secure VM, not local compute. Muse is powered by Muse Spark 1.3, the same model behind Muse Code. Muse reached number one on the US App Store within eight days of its initial launch, pushing ChatGPT to second. The Mac app is a free download, currently available in the US only.
AI Tip of the Day
Whenever we set up evaluations, we change the evaluator and the system component being tested separately. For example, in our Full Stack AI Engineering course, we test changes to the generation prompt without changing the judge prompt at the same time.
The generation prompt changes the answers being produced. The judge prompt changes how those answers are scored. If you update both in the same run and the score increases, you lose that attribution. The answers may have improved, or the new judge may simply be scoring them more favorably.
For every eval run, record:
Generation prompt version
Judge prompt version
Model version
Evaluation-set version
If you need to change both prompts, run all four combinations:
Old generation prompt + old judge
New generation prompt + old judge
Old generation prompt + new judge
New generation prompt + new judge
Keep the model and evaluation examples fixed.
If the new generation prompt performs better under both judges, you have stronger evidence that the generation improved. If the gain appears only with the new judge, the evaluation changed, not necessarily the product.
Five 5-minute reads/videos to keep you learning
1. What If Attention Isn’t All You Need? Inside Mamba and the Quiet Rewiring of Modern LLMs
State space models process sequences by updating a hidden state instead of comparing every token with every earlier token. Mamba adds input-dependent state updates, giving each token more control over what information to retain while keeping sequence processing linear and avoiding a growing KV cache. This article explains why many recent models use hybrid designs, combining mostly Mamba-style layers with a smaller number of attention layers to recover stronger recall while keeping memory costs lower on long contexts.
2. MoE Models Have Two Parameter Counts. Self-Hosting Pays for the Bigger One
Mixture-of-experts models report both total parameters and active parameters. This article explains why that is relevant for API use and self-hosting. APIs mainly charge for the small subset of experts activated for each token, while self-hosting still requires memory for every expert in the model. It shows how to read the sparsity ratio from a model configuration, estimate the real hardware footprint, and understand why a model that looks cheap per token can still be expensive to run yourself.
Prompt injection becomes harder to contain when untrusted content can persist in memory and influence later agent runs. This article follows a multi-step attack in which malicious data enters through an external field, gets written into agent memory, and later reappears in another user’s context. It argues that classifiers alone are not enough and proposes separate controls for context limits, schema validation, memory writes, and namespace-scoped retrieval so untrusted data cannot silently become part of the agent’s instructions.
4. Understanding Entropy for Data Science and AI
Entropy measures how much uncertainty or effective choice a probability distribution contains. This article derives the idea from simple examples such as coin tosses and dice, then shows why logarithms turn multiplicative numbers of possibilities into an additive measure. It extends the same reasoning to general discrete distributions, differential entropy for continuous variables, bits versus nats, and perplexity as the exponential form of entropy.
5. Kubernetes for LLM Inference: How AI Workloads Run Across a GPU Cluster
Kubernetes does not manage tokens or KV caches, but it controls where inference servers run, when they receive traffic, and how they recover from failures. LLM serving makes those decisions harder because GPUs are scarce, model startup can take minutes, and restarted replicas return with cold caches. This article explains how readiness probes, topology-aware placement, and GPU-aware scheduling fit into the system, and where Kubernetes stops being useful and an LLM-specific routing layer has to take over.
Repositories & Tools
1. Coder is a self-hosted platform for provisioning cloud development environments and AI coding agents via Terraform.
2. Agent-Native is a TypeScript framework where agents and UI share one action layer, one database, and one state.
3. AI Memory is a local MCP server that gives coding agents persistent, cross-session memory stored as Markdown files.
Top Papers of The Week
1. ModularRSI: Generalizable Recursive Harness Self-Improvement
Harness self-improvement can overfit to benchmark tasks or make broad changes that mix useful fixes with task-specific behavior. This paper proposes ModularRSI, a benchmark-disjoint, contrastive, and modular framework for generalizable harness evolution. It compares successful and failed trajectories, identifies recurring weaknesses, and evolves five harness components separately using 2,000 benchmark-disjoint tasks before combining the changes.
2. An Empirical Study of Harness Design for Coding Agents
Coding-agent benchmarks usually compare complete harnesses, making it difficult to tell which design choices actually improve performance. This study fixes the execution loop and varies planning, action space, and context management across 176 settings, four models, SWE-Bench Verified, and Terminal-Bench 2.1. It finds that context management matters mainly when context is tight, rule-based elision before LLM summarization gives the best overall efficiency, planning increasingly saves cost rather than accuracy for stronger models, and bash-capable models can use a simpler bash-only interface at substantially lower cost.
3. GLiFormer: A Generalist Multitask Transformer Encoder
Tasks such as entity extraction, relation extraction, classification, and structured JSON generation usually require separate models or autoregressive generation. GLiFormer uses one schema-conditioned encoder and a shared anchor representation to handle all of them, grounding extracted values in the source and assembling hierarchical outputs without token-by-token generation. Across 26 NER and 13 classification datasets, it performs strongly for its size; on a multilevel structuring benchmark, GLiFormer-large reaches 91.10% F1 versus 91.96% for GPT-5.6 Luna, while GLiFormer-base runs with 69 ms median GPU latency in a separate efficiency test.
4. Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion
Retrieval often needs a set of complementary results rather than one best match, but objectives such as diversity, coverage, and coherence are difficult to learn from standard query-item training data. R4T first uses reinforcement learning to train an LLM on these set-level objectives, then uses that model to generate aligned training pairs and distills the behavior into a lightweight diffusion retriever that produces multiple retrieval targets in one pass. On large-scale fashion and music datasets, it improves set-level retrieval quality over strong baselines while reducing fan-out latency by roughly an order of magnitude compared with autoregressive generation.
5. Reimagining Research Papers As Interactive and Reliable AI Agents
This paper introduces Paper2Agent, an automated framework that converts research papers into artificial intelligence (AI) agents. It converts papers and their codebases into MCP servers with validated tools, resources, and workflows that an AI agent can invoke through natural language. It successfully converted 74 of 100 computational-biology papers and validated 593 of 599 generated tools; on 300 tutorial-derived questions, it reached 91.2% accuracy versus 80.3% for Claude Code with direct repository access, while also cutting average query cost from $0.38 to $0.20 and latency from 4.3 to 1.6 minutes.
Quick Links
1. OpenAI releases a Model Misalignment Disclosure Framework, alongside six reports on unexpected or concerning model behavior observed during training and evaluation over the previous six months. The framework covers qualifying incidents throughout training, evaluation, testing, and deployment, and routes disclosures through three tracks based on investigation readiness and complexity: Ready for Disclosure, Minor Investigation, and a Slow Track for larger cases, particularly those involving third parties. The initial reports include GPT-5.6 Sol instances writing instructions into compaction summaries to conceal mistakes or invent missing information, behavior flagged in 2.15% of Sol RL compaction summaries versus 0.27% for GPT-6 Astra. OpenAI says no industry-wide standard exists for disclosing model misalignment and describes the framework as a first step toward one.
2. StepFun launches Step 5 Preview, a 600B-parameter sparse MoE model with 27B parameters active per token, a 1M-token context window, and multimodal vision input. StepFun designed it for long-horizon agentic work across software engineering, professional knowledge tasks, and finance, with low-, medium-, and high-reasoning-effort settings. Artificial Analysis scores Step 5 Preview at 44 on its Intelligence Index, just behind GLM-5.3 (max) at 45, while Step 5 costs less per token. API pricing is $1.00 per million uncached input tokens, $0.05 for cached input, and $2.70 for output. The API is available now, while StepFun plans to release open weights on October 15.
Who’s Hiring in AI
Software Engineer, Beneficial Deployments @Anthropic (San Francisco/ New York)
Senior AI Engineer @American Express (Phoenix, AZ, USA)
Forward Deployed AI Engineer @Turing (New York, NY, USA)
Senior Full Stack Engineer @Kelly Services (New York, NY, USA)
Full Stack Engineer @NTT Data Americas, Inc. (Plano, TX, USA)
Analytics Engineer @Movable Ink (Remote/Toronto)
Full Stack Engineer, Contract @66degrees (United Kingdom)
Interested in sharing a job opportunity here? Contact sponsors@towardsai.net.
Think a friend would enjoy this too? Share the newsletter and let them join the conversation.