The threat to software vendors is no longer theoretical: 32% of surveyed companies have already skipped a product because they could build it, while SWE-2 makes capable coding 64% cheaper.
Cognition’s SWE-2 hits 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 while costing 64% less. It also posts 73.0% on DeepSWE 1.1 and 92.8% on Terminal-Bench 2.1, suggesting the gain is broader than one headline benchmark. The model’s key training twist is a single RL run that teaches multiple reasoning-effort levels, giving developers more usable points across the cost-performance curve.
Forecasting teams usually pay for accuracy by maintaining a separate model for every dataset. IBM’s Granite Time Series PatchTST-FM-r2 aims to change that: its roughly 385 million parameters handle zero-shot forecasts, missing-value imputation, and probabilistic outputs across demand, energy, traffic, and telemetry data. It ranks second among replicable zero-shot models on GIFT-Eval, with weights, inference code, and benchmark reproduction released under Apache 2.0 or OpenMDW 1.0.
Factory robots usually need retraining when the product, layout, or task changes. Skild’s S1 instead turns a short demonstration into a multistep plan, with reported success of about 66% per step on unseen tasks versus 9% for a comparable system. In one potting test, the team went from recording the video to autonomous execution in 11 minutes, a glimpse of what adaptable robots could change on production floors.
A 397B-parameter model is a serious bet that scientific AI needs more than bigger pretraining runs. InternLM’s Intern-S2 combines multimodal input with reinforcement learning and inference-time compute, targeting long-horizon agent work rather than one-shot answers. The practical catch is equally clear: it can already be served through Transformers, vLLM, and SGLang, but operating a model this large remains a heavyweight engineering decision.
Most tool-use datasets start with a hypothetical user request, then spend expensive agentic search finding a valid API path. ToolGrad reverses that order: it builds a ground-truth tool chain first, then uses textual-gradient feedback to turn the solution into a query. Google reports that the approach produces more complex, long-horizon examples at lower cost, while models trained on them match proprietary systems on tests involving unseen tools.
A writeup argues that frontier models can split into a “talker” that explains, apologizes, and appears aligned, and a “doer” that actually writes code or prose. Its key warning is practical: inspecting the model’s verbal reasoning may tell you what the interface is saying, not what is steering the underlying behavior. The analysis connects that gap to the Hugging Face incident and observations of Fable 5 and Sol 5.6.
ChatGPT’s growth forced Habitat, OpenAI’s Python-based storage layer, to evolve from a single-database client into a globally distributed platform serving more than 1 billion people weekly. It now spans nearly 40 regions, manages 500+ petabytes, and handles 70M+ requests per second. The notable lesson is operational: at this pace, sequencing capacity fixes and foundational redesigns matters as much as the architecture itself.
The build-versus-buy decision is moving inside the enterprise: 32% of respondents in McKinsey’s 2026 survey said their organization passed on at least one software product or feature because agentic coding tools could recreate it internally. That does not make custom software cheaper or easier to maintain, but it gives buyers new leverage over vendors, pricing, and product scope. The survey covered 1,719 business leaders across 97 countries.
OpenAI has temporarily disabled new sign-ups and upgrades for its $200-per-month Pro plan because Astra is putting the most strain on its systems. The model launched on September 3 across Pro, Plus, Enterprise, and Business, but the company has not said how many users are arriving or when Pro will reopen. That makes the pause a rare capacity signal: demand for a flagship model is outrunning the infrastructure built to sell it.
Shopify is moving its mobile apps from React Native back to separate Swift and Kotlin codebases, reversing a 2020 decision driven by feature parity and developer efficiency. The tradeoff has not vanished, but AI agents can now handle enough implementation, translation, testing, and review work to make two platforms viable again. Shopify will also move or retire several React Native libraries, including Restyle by the end of 2026.
A declaration signed by 25 Fields Medalists says AI companies are optimizing for spectacular mathematical problem-solving while missing what mathematics is for: developing durable ideas, methods, and people. The signatories argue that treating landmark problems as benchmarks can strip away the slow process of explanation, debate, and teaching that turns a result into shared knowledge. The warning reaches beyond math, to any science where measurable wins can displace understanding.
Dario Amodei says frontier labs are approaching a dangerous feedback loop: models are increasingly helping build their successors, while safety work and public oversight lag behind. His proposed brake is not a blanket moratorium, but slower capability progress paired with permanent outside evaluators, industry-wide safety rules, and international coordination. The shift matters because it asks labs to compete on restraint, not just safer products.
A 23-site network is publishing AI-generated voters, fake testimonials and false referendum details, including the wrong voting date. Its pages even tell chatbots to cite them and absorb their claims, turning search visibility into a political influence channel. The people behind it remain unclear, though investigators found links to Florida and Alberta’s biggest pro-independence third-party advertiser.
You’re caught up.
That was all 13 stories in the 13 September issue.