Astra directly edited a small fraction of Qwen3.5-0.8B's weights, raising its MATH-500 score from 7.9% to 48.1% under identical sampling settings. SemiAnalysis released GDPx, covering 206 tasks across 11 economic sectors. Even the top scorer, Opus 5.5, reached only 12%, while Gemini 4 Carbon scored 10.9% at roughly a quarter of the cost. Oracle plans to raise $20 billion through common stock and senior notes, with proceeds available for more than AI infrastructure expansion, yet its shares still fell 8.8% that day.
Astra turned the reasoning-loop problem flagged in the model card into a target for weight-level repair.
Researchers demonstrated “weight-space programming”: without gradient descent, Astra directly edited 0.016% of Qwen3.5-0.8B's weights. Under identical sampling settings, its MATH-500 score rose from 7.9% to 48.1%, while its score on the IFEval instruction-following benchmark tripled. The changes targeted the reasoning-loop problem already flagged in the model card, demonstrating a way to debug models by directly correcting their weights.
Read original →GDPx measures tasks across industries, while Legora tests legal agents through simulated client work.
SemiAnalysis released the GDPx benchmark, covering 206 agent tasks across 11 economic sectors, including chip design, software engineering, finance, and biology. Even the best-performing model, Opus 5.5, scored just 12%. Gemini 4 Carbon followed closely at 10.9%, at roughly a quarter of the cost, illustrating different score-versus-cost trade-offs.
Read original →Legora's CTO described how a team of three engineers and three lawyers built an evaluation environment that simulates client work—for example, asking agents to conduct due diligence in a data room containing 10,000 documents and produce a report identifying key risks. Unlike teams that improve agents by examining execution traces, Legora cannot use that approach because of client confidentiality. The team maps evaluation results into heatmaps by legal practice area and task, because a new model may improve on some tasks while regressing on others.
Read original →🔍 Analysis: The score-versus-cost trade-offs between Opus 5.5 and Gemini 4 Carbon on GDPx still need to be tested against heatmaps broken down by practice area in legal tasks like Legora's: a higher overall score does not mean better performance in every area. Client confidentiality also prevents Legora from relying on real execution traces to improve its agents. This means enterprise model selection requires repeatable, simulated work environments, rather than model upgrades based solely on general-purpose leaderboards.
DeepSeek constrains when agents act, while OpenAI builds a backend for the development scale driven by Codex.
DeepSeek released dsh-superagent, a research prototype that works on a single GitHub repository, handles tasks in separate sessions and directories, and returns results such as PRs. It acts only when triggered by an authorized request or team rules, and speaks only when someone needs to take action or can use the result, reducing demands on team attention. The prototype is neither a formal product nor production-ready; the public version differs from the internal implementation, and some design elements have not yet been implemented.
Read original →OpenAI published “Scaling Git at OpenAI,” introducing the Pando backend it has built. Pando is Git-compatible and designed to support the scale of development driven by Codex. OpenAI's Thibault Sottiaux said the team could finally discuss the work publicly.
Read original →🔍 Analysis: dsh-superagent and Pando address two kinds of capacity constraints that arise when coding agents join teams. The former manages authorization, communication, and task isolation to avoid draining human attention; the latter supports the scale of Git-based development driven by Codex. For enterprises, adopting coding agents is not just about producing more code: it also means testing whether collaboration rules and repository infrastructure can handle the load. DeepSeek's prototype is not yet production-ready and cannot be treated as a deployable solution.
Oracle seeks expansion funding, while NVIDIA's AgentX submission emphasizes inference throughput and total cost of ownership.
On November 2, Oracle filed plans to issue $10 billion each in common stock and senior notes to fund AI infrastructure expansion. The financing had been previewed at the previous week's investor day, but the company's shares still fell 8.8% that day. Oracle said $20 billion is the ceiling for additional capital-market financing during its fiscal 2027–2030 investment cycle. Proceeds may also be used to repay debt and fund future investments, rather than solely for expansion.
Read original →NVIDIA's AgentX submission showcased inference software written by a swarm of agents. On a 700B+ model, the software can deliver more than 1,000 tokens per second per user. The submission also claimed that, at 325 tokens per second per user, Rubin NVL72 achieves a performance-to-total-cost-of-ownership ratio 20.9 times that of GB300 NVL72.
Read original →🔍 Analysis: Oracle's $20 billion financing figure measures capital input, while AgentX's claimed 20.9-fold performance-to-total-cost-of-ownership ratio measures hardware efficiency under specific throughput conditions. Neither translates directly into an equivalent increase in business returns. For buyers, cost comparisons at more than 1,000 tokens per second per user and at 325 tokens per second per user also use different test conditions. Assessing infrastructure economics requires considering throughput conditions alongside costs.
JEPA research examines alignment signals, while FineDNA releases the data and recipes used to train a DNA model.
Research by Balestriero and LeCun reports that, under specific conditions, independently trained image and text JEPAs converge to similar representations even without paired data. Their image-text alignment used 21.6 million samples, compared with 900 million for next-token prediction—roughly 1/42 as many. The study attributes the result to the training signal of “two views of the same thing,” rather than scale.
Read original →Hugging Face released FineDNA, a mid-training dataset totaling 514.3 billion tokens across genomes, transcripts, biological text, and general text. It was used to train Qwen3-14B into Carbon-X, a model that can read DNA; the actual mid-training run consumed 302 billion tokens. The dataset covers all three training stages and includes the recipes for each, along with sampling weights for each data source in mid-training.
Read original →