Uncategorized

In this tutorial, we perform an advanced single-cell RNA-seq analysis workflow using Scanpy on the PBMC-3k benchmark dataset. We start by loading the dataset, inspecting its structure, and applying quality control checks to evaluate gene counts, total counts, mitochondrial content, and ribosomal gene signals. We then filter low-quality cells and genes, detect potential doublets with Scrublet, normalize the data, apply log transformation, and identify highly variable genes for downstream analysis. Also, we score cell-cycle phases, regress out unwanted technical variation, scale the data, and reduce dimensionality using PCA, UMAP, and t-SNE. We also cluster cells with the Leiden algorithm, identify marker genes, annotate cell populations using canonical PBMC markers, explore trajectory structure with PAGA and diffusion pseudotime, calculate a custom interferon-response score, and finally save the fully analyzed AnnData object for future use. Copy CodeCopiedUse a different Browser !pip install -q scanpy leidenalg python-igraph scrublet import scanpy as sc import numpy as np import pandas as pd import matplotlib.pyplot as plt import warnings warnings.filterwarnings(“ignore”) sc.settings.verbosity = 3 sc.settings.set_figure_params(dpi=80, facecolor=”white”, figsize=(5, 5)) sc.logging.print_header() adata = sc.datasets.pbmc3k() adata.var_names_make_unique() print(adata) adata.var[“mt”] = adata.var_names.str.startswith(“MT-“) adata.var[“ribo”] = adata.var_names.str.startswith((“RPS”, “RPL”)) sc.pp.calculate_qc_metrics( adata, qc_vars=[“mt”, “ribo”], percent_top=None, log1p=False, inplace=True ) sc.pl.violin( adata, [“n_genes_by_counts”, “total_counts”, “pct_counts_mt”], jitter=0.4, multi_panel=True, ) sc.pl.scatter(adata, x=”total_counts”, y=”pct_counts_mt”) sc.pl.scatter(adata, x=”total_counts”, y=”n_genes_by_counts”) We install the required single-cell analysis libraries and import Scanpy, NumPy, Pandas, Matplotlib, and warning controls. We load the PBMC-3k benchmark dataset, make gene names unique, and inspect the AnnData object structure. We then calculate quality control metrics for mitochondrial and ribosomal genes and visualize count-level quality patterns using violin and scatter plots. Copy CodeCopiedUse a different Browser sc.pp.filter_cells(adata, min_genes=200) sc.pp.filter_genes(adata, min_cells=3) adata = adata[adata.obs.n_genes_by_counts < 2500, :].copy() adata = adata[adata.obs.pct_counts_mt < 5, :].copy() sc.pp.scrublet(adata) print(“Predicted doublets:”, int(adata.obs[“predicted_doublet”].sum())) adata = adata[~adata.obs[“predicted_doublet”], :].copy() adata.layers[“counts”] = adata.X.copy() sc.pp.normalize_total(adata, target_sum=1e4) sc.pp.log1p(adata) sc.pp.highly_variable_genes(adata, min_mean=0.0125, max_mean=3, min_disp=0.5) sc.pl.highly_variable_genes(adata) adata.raw = adata adata = adata[:, adata.var.highly_variable].copy() We filter out low-quality cells and rarely detected genes to improve the reliability of the dataset. We use Scrublet through Scanpy to identify predicted doublets and remove them before deeper analysis. We then preserve raw counts, normalize expression values, apply log transformation, select highly variable genes, and keep only the most informative features. Copy CodeCopiedUse a different Browser s_genes = [“MCM5″,”PCNA”,”TYMS”,”FEN1″,”MCM2″,”MCM4″,”RRM1″,”UNG”,”GINS2″, “MCM6″,”CDCA7″,”DTL”,”PRIM1″,”UHRF1″,”HELLS”,”RFC2″,”NASP”, “RAD51AP1″,”GMNN”,”WDR76″,”SLBP”,”CCNE2″,”UBR7″,”POLD3″,”MSH2″, “ATAD2″,”RAD51″,”RRM2″,”CDC45″,”CDC6″,”EXO1″,”TIPIN”,”DSCC1″, “BLM”,”CASP8AP2″,”USP1″,”CLSPN”,”POLA1″,”CHAF1B”,”E2F8″] g2m_genes = [“HMGB2″,”CDK1″,”NUSAP1″,”UBE2C”,”BIRC5″,”TPX2″,”TOP2A”,”NDC80″, “CKS2″,”NUF2″,”CKS1B”,”MKI67″,”TMPO”,”CENPF”,”TACC3″,”SMC4″, “CCNB2″,”CKAP2L”,”CKAP2″,”AURKB”,”BUB1″,”KIF11″,”ANP32E”, “TUBB4B”,”GTSE1″,”KIF20B”,”HJURP”,”CDCA3″,”CDC20″,”TTK”, “CDC25C”,”KIF2C”,”RANGAP1″,”NCAPD2″,”DLGAP5″,”CDCA2″,”CDCA8″, “ECT2″,”KIF23″,”HMMR”,”AURKA”,”PSRC1″,”ANLN”,”LBR”,”CKAP5″, “CENPE”,”NEK2″,”G2E3″,”CBX5″,”CENPA”] s_genes = [g for g in s_genes if g in adata.var_names] g2m_genes = [g for g in g2m_genes if g in adata.var_names] sc.tl.score_genes_cell_cycle(adata, s_genes=s_genes, g2m_genes=g2m_genes) sc.pp.regress_out(adata, [“total_counts”, “pct_counts_mt”]) sc.pp.scale(adata, max_value=10) sc.tl.pca(adata, svd_solver=”arpack”) sc.pl.pca_variance_ratio(adata, log=True, n_pcs=50) sc.pp.neighbors(adata, n_neighbors=10, n_pcs=40) sc.tl.umap(adata) sc.tl.tsne(adata, n_pcs=40) We define S-phase and G2/M-phase marker genes and retain only those present in the dataset. We score each cell for cell-cycle phase, regress out unwanted variation from total counts and mitochondrial percentage, and scale the data for downstream modeling. We then run PCA, inspect explained variance, construct the neighborhood graph, and generate UMAP and t-SNE embeddings. Copy CodeCopiedUse a different Browser sc.tl.leiden(adata, resolution=0.5, flavor=”igraph”, n_iterations=2, directed=False) sc.pl.umap(adata, color=”leiden”, legend_loc=”on data”, title=”Leiden clusters”) sc.pl.tsne(adata, color=”leiden”, legend_loc=”on data”, title=”t-SNE clusters”) sc.tl.rank_genes_groups(adata, “leiden”, method=”wilcoxon”) sc.pl.rank_genes_groups(adata, n_genes=20, sharey=False) result = adata.uns[“rank_genes_groups”] groups = result[“names”].dtype.names top_df = pd.DataFrame({g: result[“names”][g][:10] for g in groups}) print(“nTop 10 markers per cluster:n”, top_df) marker_genes = { “B-cell”: [“CD79A”, “MS4A1”], “CD8 T-cell”: [“CD8A”, “CD8B”], “CD4 T-cell”: [“IL7R”, “CD4”], “NK”: [“GNLY”, “NKG7”], “CD14 Monocyte”: [“CD14”, “LYZ”], “FCGR3A Monocyte”: [“FCGR3A”, “MS4A7”], “Dendritic”: [“FCER1A”, “CST3”], “Megakaryocyte”: [“PPBP”], } sc.pl.dotplot(adata, marker_genes, groupby=”leiden”, standard_scale=”var”) sc.pl.stacked_violin(adata, marker_genes, groupby=”leiden”, swap_axes=True) We apply Leiden clustering to group cells based on the neighborhood graph and visualize the clusters on UMAP and t-SNE plots. We perform differential expression analysis using the Wilcoxon test to identify the top marker genes for each cluster. We then use canonical PBMC marker genes to support cell-type annotation through dot plots and stacked violin plots. Copy CodeCopiedUse a different Browser sc.tl.paga(adata, groups=”leiden”) sc.pl.paga(adata, color=”leiden”, threshold=0.1) sc.tl.umap(adata, init_pos=”paga”) sc.pl.umap(adata, color=”leiden”, legend_loc=”on data”) sc.tl.diffmap(adata) sc.pp.neighbors(adata, n_neighbors=10, use_rep=”X_diffmap”) adata.uns[“iroot”] = np.flatnonzero(adata.obs[“leiden”] == adata.obs[“leiden”].cat.categories[0])[0] sc.tl.dpt(adata) sc.pl.umap(adata, color=[“leiden”, “dpt_pseudotime”], legend_loc=”on data”) ifn_genes = [“ISG15”, “IFI6”, “IFIT1”, “IFIT3”, “MX1”, “OAS1”, “STAT1”, “IRF7″] ifn_genes = [g for g in ifn_genes if g in adata.raw.var_names] sc.tl.score_genes(adata, gene_list=ifn_genes, score_name=”IFN_score”) sc.pl.umap(adata, color=”IFN_score”, cmap=”viridis”) adata.write(“pbmc3k_analyzed.h5ad”) print(“n Analysis complete — saved to pbmc3k_analyzed.h5ad”) print(adata) We run PAGA to model connectivity between Leiden clusters and reinitialize UMAP using the PAGA graph to obtain a clearer trajectory structure. We compute diffusion maps and diffusion pseudotime to explore possible progression patterns across cell states. We also calculate an interferon-response gene-set score, visualize it on UMAP, and save the final analyzed object as an .h5ad file. In conclusion, we built an end-to-end Scanpy pipeline for single-cell RNA-seq analysis, transforming raw PBMC data into interpretable biological insights. We cleaned and preprocessed the dataset, removed noisy cells and doublets, selected informative genes, and generated meaningful embeddings to visualize cellular structure. We then used Leiden clustering and differential expression analysis to discover marker genes and connect clusters to known immune cell types. By adding PAGA, diffusion pseudotime, and custom gene-set scoring, we extended the workflow beyond basic clustering and showed how Scanpy supports deeper biological interpretation. At last, we had a saved .h5ad object that contains the processed data, annotations, scores, clusters, and visual analysis results, ready for downstream exploration or reporting. Check out the Full Codes with Notebook here. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post How to Build a Single-Cell RNA-seq Analysis Pipeline with Scanpy for PBMC Clustering, Annotation, and Trajectory Discovery appeared first on MarkTechPost.

MIT Technology Review Explains: Let our writers untangle the complex, messy world of technology to help you understand what’s coming next. You can read more from the series here. Eight passengers aboard a Dutch-flagged cruise ship have contracted a type of hantavirus, a rare virus transmitted by rats. Three of them have died. As the ship prepares to dock in the Canary Islands, plans are being finalized to let the remaining passengers and crew disembark safely. The virus in question appears to have a high fatality rate. Read on for answers to the big questions surrounding the outbreak—and to hear why health experts don’t expect a rerun of the covid-19 pandemic. What is hantavirus? Hantaviruses are a group of viruses that typically infect rodents but can be transmitted to humans through exposure to the animals or their droppings, urine, or saliva. The viruses don’t seem to cause illness in rodents, but they can make people very unwell. The symptoms can depend on the type of hantavirus a person has been exposed to. Varieties found in the Americas can cause hantavirus cardiopulmonary syndrome, which affects the lungs and heart and has a fatality rate of up to 50%. That condition made headlines last year when it caused the death of pianist Betsy Arakawa, the wife of actor Gene Hackman.  How many cases have there been so far? On April 6, a man aboard the MV Hondius developed respiratory symptoms. He became very unwell and died just five days later. His wife, who left the ship at the island of Saint Helena, also developed symptoms. Her health deteriorated during a flight to Johannesburg, South Africa, and she died the following day, on April 26. South Africa’s National Institute of Communicable Diseases tested samples taken from the woman and confirmed that she had hantavirus. A third person aboard the ship, who developed symptoms on April 28, died on May 2. Four other passengers who became ill were evacuated—one to South Africa and three to the Netherlands. An eighth person had disembarked in Saint Helena and reported similar symptoms once he was in Zurich, Switzerland. A team at Geneva University Hospitals confirmed that he had become ill from the Andes virus—a form of hantavirus that can be spread between people. Could this be the start of the next pandemic? Health experts don’t believe so. They stress that the situation is nothing like the one the coronavirus that causes covid-19 presented in 2020. For a start, the Andes virus is not a mysterious new virus—scientists already have an understanding of it, and Argentina is sharing diagnostic kits it has already developed. The virus also doesn’t spread in the same way. Officials at the World Health Organization emphasized that the spread of hantavirus requires close contact—the kind a person might have with a partner, household member, or medical caregiver. The cruise ship outbreak represents “a specific confined setting where people are interacting in a prolonged close contact,” Abdirahman Mahamud, the alert and response director for the WHO’s health emergency program, said at a press event on Thursday. “With the experience our member states have, and the actions they have taken, we believe that this will not lead to a subsequent chain of transmission.” What about the rest of the people onboard the ship? All the remaining passengers have been asked to stay in their cabins, which the WHO says are being disinfected. Doctors and health professionals from the WHO and the European Center for Disease Prevention and Control have boarded the ship and are assessing everyone on board. So far, no one else on board has developed symptoms, Maria Van Kerkhove, WHO acting director for epidemic and pandemic management, said at the press event. That’s “a good sign,” she said, but she added that the Andes virus has a long incubation period (around six weeks). Passengers are being advised to wear a medical mask when they leave their rooms. At the same event, WHO director general Tedros Adhanom Ghebreyesus said he was in regular contact with the ship’s captain, who was reporting that “morale had increased significantly” since the ship started its journey to the Canary Islands. What do we know about the Andes virus? The Andes virus is the only hantavirus that is known to be transmitted between people. That transmission seems to rely on prolonged, intimate contact. There was an Andes virus outbreak in Argentina around eight years ago. Between November 2018 and February 2019, there were 34 confirmed cases of infection, and 11 deaths. That outbreak was triggered when a person with symptoms attended a social gathering, said Tedros. “We are in a similar situation right now,” he said. “A cluster in a confined space with close contact.” The fact that the 2018 outbreak was limited to 34 cases should be somewhat reassuring, he implied. “We believe this will be a limited outbreak if the public health measures are implemented and solidarity is shown across all countries,” he said. How is hantavirus treated? Unfortunately, we don’t have any specific antiviral treatments or vaccines for hantavirus. The WHO recommends early intensive care for people who develop symptoms. “This can save lives,” Anaïs Legand, WHO technical lead on viral hemorrhagic fevers, said on Thursday. How did people get infected in the first place? We don’t yet have an answer to that. But we do know that the couple who died had traveled through Argentina, Chile, and Uruguay on a birdwatching trip before they boarded the ship. That trip included visits to areas where species of rats that carry the Andes virus are known to live. The WHO is working with authorities in Argentina to try to retrace the couple’s movements on that trip. Has the virus spread beyond the ship? We don’t yet know for sure. The WHO is receiving reports of “potential suspect cases,” Van Kerkhove said at the Thursday briefing. Some of them have links to the ship or its passengers. Each “alert” will be followed up by health authorities in the relevant

OpenAI has launched a Codex Chrome extension for Mac and PC to streamline browser-based workflows that were previously difficult to handle via APIs or plugins. This release follows a trend where most users preferred working in a browser after the launch of “Computer Use,” allowing Codex to operate more effectively across various web-based tasks. What the Extension Actually Does Before this release, Codex had access to an in-app browser — a sandboxed browser built into the Codex desktop app itself — and a growing library of dedicated plugins for services like GitHub, Slack, Figma, and Notion. The new Chrome extension fills a gap those two approaches couldn’t cover: tasks that require your real, signed-in browser state. The Codex Chrome extension lets Codex use Chrome for browser tasks that need your signed-in browser state. It is intended for use when Codex needs to read or act on sites such as LinkedIn, Salesforce, Gmail, or internal tools. For everything else like local development servers, file-backed previews, and public pages that do not require a sign-in you can continue using the in-app browser, which keeps that preview and verification work inside Codex without touching your Chrome profile. Codex now operates across three distinct tool tiers depending on the task: plugins when a dedicated integration is available, Chrome when it needs logged-in browser context, and the in-app browser for localhost. The agent selects which tier to use automatically, though users can also invoke Chrome directly in a prompt using the @Chrome mention syntax — for example: @Chrome open Salesforce and update the account from these call notes. If Chrome isn’t already open, Codex can open it. On the functional side, the new browser-based capabilities of the plugin include testing web apps, collecting context from across open tabs, and using Chrome DevTools in parallel while the user performs other tasks. Critically, Codex works in task-specific tab groups, so it can gather context and take actions without taking over your active browsing session. How to Install and Use Codex Chrome extension Quick Start Guide Installing and Using the Codex Chrome Extension Five steps to connect Codex to your signed-in browser. Works on macOS and Windows. Not available in EU or UK yet. 1 Install the extension from the Chrome Web Store Open Chrome and go to the Codex listing in the Chrome Web Store. Click Add to Desktop and confirm the prompt that appears. Chrome Web Store listing for Codex by OpenAI Shows the Codex extension card with version, publisher, and Add to Desktop button </> Codex by OpenAI · v1.1.4 · 109 KiB · Tools Control Chrome with Codex Add to Desktop Codex does not support other Chromium-based browsers (Brave, Edge, Arc) at this time. 2 Add the Chrome plugin inside the Codex app Open the Codex desktop app and navigate to Plugins. Find the Chrome plugin and click Add. Codex will walk you through the connection flow. Codex App › Plugins › Chrome › Add Chrome Let Codex use your signed-in browser Add 3 Approve Chrome permissions and confirm “Connected” Chrome will ask you to accept a set of extension permissions. After approving, open Chrome and confirm the Codex extension shows Connected in the toolbar. Permissions requested by the extension Access the page debugger Read and change all your data on all websites Read and change your browsing history on all signed-in devices View and manage your tab groups Manage your downloads Extension status in Chrome toolbar Connected These permissions let the extension operate browser workflows. Codex still applies its own per-site confirmation prompts and allowlist/blocklist on top of these Chrome permissions. 4 Start a new Codex thread and invoke Chrome Open a new thread in Codex. You can let Codex choose the right tool automatically, or invoke Chrome directly using the @Chrome mention. Codex will open Chrome if it isn’t already running. @Chrome open Salesforce and update the account from these call notes. You can also just describe the task naturally — Codex will select Chrome automatically when the task requires a signed-in website. 5 Review and approve site access when prompted By default, Codex asks before interacting with each new website host. Choose one of three options each time it asks. Manage a permanent allowlist and blocklist in Computer Use settings. Codex is asking to use salesforce.com Allow this chat Always allow host Decline Example: Codex working in Chrome Codex · New thread @Chrome Go to my LinkedIn notifications, summarize any messages from recruiters, and draft a short reply to each. Opening Chrome · Requesting access to linkedin.com Working in tab group · Your active tabs are unaffected Found 3 recruiter messages in your LinkedIn inbox. Here’s a summary and a draft reply for each: Sara H. — Eng Manager role Draft ready James K. — ML Lead, Series B Draft ready Priya M. — Staff Engineer Draft ready Task complete · Pages kept open for review Codex tab group in Chrome LinkedIn — Inbox Message · Sara H. +2 more Before you use it — three things to know Treat page content as untrusted. Malicious pages can embed instructions designed to redirect Codex’s behavior (prompt injection). Review websites before approving access. Memories setting affects browser tasks. If Memories is on, Codex can use previously stored context during Chrome tasks. Turn Memories off for fully isolated sessions. File uploads need an extra setting. Go to Chrome Extensions › Codex › Details and enable Allow access to file URLs if any task requires uploading a local file. Permission Model and Security Considerations The Codex Chrome extension requires extensive browser permissions—including access to your history, bookmarks, and page data—to function as an integrated AI agent. To mitigate security risks, OpenAI has implemented a per-site confirmation layer that asks for permission before Codex interacts with a new domain. Users can manage these permissions via an allowlist or blocklist, though certain features like browser history access carry higher risk and require manual approval for each session. Beyond standard permissions, developers must be wary of prompt

In the second week of the landmark trial between Elon Musk and OpenAI, Musk’s motivations for bringing the suit were under scrutiny. Last week, Musk took the stand, alleging that OpenAI CEO Sam Altman and president Greg Brockman had deceived him into donating $38 million to the company. He claimed that they’d promised to maintain it as a nonprofit dedicated to developing AI for the benefit of humanity, only to later accept billions of dollars of investment from Microsoft and restructure the company to operate a for-profit subsidiary.   This week, Brockman fired back with his side of the story, arguing that Musk had actually pushed for OpenAI to create a for-profit arm and fought a bitter battle to have “absolute control” over it. OpenAI has argued that Musk is suing because he didn’t get his way and is now trying to undermine a competitor to his own AI company, xAI. Shivon Zilis, a former OpenAI board member and the mother of four of Musk’s children, also testified, revealing that Musk tried to recruit OpenAI CEO Sam Altman to lead a new AI lab at his electric-car company, Tesla.  Musk cofounded OpenAI in 2015 with Altman, Brockman, and others but left in 2018. Now, he’s asking the court to remove Altman and Brockman from their roles and to unwind the restructuring OpenAI undertook last year, which converted its for-profit subsidiary into a public benefit corporation. He is also seeking as much as $134 billion in damages from OpenAI and Microsoft, OpenAI’s investor.  The outcome of the trial could upend OpenAI’s race toward an IPO at a valuation approaching $1 trillion. Meanwhile, xAI, which Musk founded in 2023, is now a division of his rocket company, SpaceX; the combined companies are also expected to go public as early as June, at a target valuation of $1.75 trillion. On Monday, Brockman walked into the courtroom in a blue suit and tie, holding hands with his wife, Anna Brockman. On the stand, he was serene, even chipper, as he recalled OpenAI’s early days. But he grew agitated under impassioned questioning from Elon Musk’s lawyer, Steven Molo. Altman listened in silence, while Anna Brockman sat behind him, fidgeting. Outside the courthouse, protesters rallying against the AI race sang hymns over the voices of lawyers giving press conferences. Two days before trial began, according to Brockman, Musk messaged him to ask if he would be interested in settling. When Brockman suggested that both sides drop their claims, Musk texted back: “By the end of this week, you and Sam will be the most hated men in America. If you insist, so it will be.” Musk stormed out with a Tesla painting Last week, Musk testified that he’s suing to save OpenAI’s nonprofit mission to develop AI safely, but he said he was open to seeing OpenAI become a capped-profit company with moderate investments from Microsoft.  This week, Brockman told the jury that Musk was never truly committed to keeping OpenAI a nonprofit. In the summer of 2017, when an AI model that OpenAI built beat the world’s best players in a video game called Dota 2, Musk hosted a gathering at his “Haunted Mansion” near San Francisco. The house was splattered with confetti and cups, Brockman recalled, and the actress Amber Heard, who was Musk’s girlfriend at the time, served whiskey. “Time to make the next step for OpenAI. This is the triggering event,” Musk wrote in an email—having said weeks earlier that if OpenAI made a major public achievement, it would be “time to create a for-profit,” Brockman told the jury. Over the next six weeks, Brockman said, Musk and the other cofounders had intense discussions about creating a for-profit entity to raise enough capital to build artificial general intelligence—powerful AI that can compete with humans on most cognitive tasks. Musk wanted to have majority equity in the entity and the right to choose a majority of the board members. He also wanted to be its CEO, said Brockman.  Brockman testified that in August 2017, he and other cofounders gathered to hash out the terms of the for-profit structure. Ilya Sutskever, OpenAI’s chief scientist at the time, arrived bearing a painting of a Tesla as a “token of goodwill” in return for the actual Teslas Musk had given them days earlier. “It felt a little bit like [Musk] was buttering us up, right,that he wanted us to feel indebted to him,” Brockman told the jury. When Brockman and Sutskever proposed that they all have equal shares of equity, said Brockman, Musk fell silent and finally said, “I decline.” Musk then stood up and “stormed around the table,” he said. “I actually thought he was going to hit me.” Musk grabbed the painting and walked out.  Brockman said that afterwards he struggled to decide whether to continue building OpenAI with Musk or break away. “There was a fork in the road,” he said. “Do we accept Elon’s terms? Or do we reject the terms, he quits to create his own, and then we create our own?” “The one thing we could not accept was to hand him unilateral, absolute control, potentially, over the AGI,” Brockman told the jury. What was Brockman thinking? In his theatrical baritone, Molo argued that Brockman was motivated by greed rather than a commitment to OpenAI’s nonprofit mission to develop AI that benefits humanity. He noted that while Brockman never invested money in the company, he now owns a stake worth close to $30 billion.  “Solving for the mission has always been my primary motivation,” Brockman said, pushing back on Molo’s characterization of him. “It remains so today.”  Molo pulled up Brockman’s electronic journal on a screen in the courtroom, trying to show the jury what Brockman was really thinking behind the scenes. In 2017, while negotiating with Musk about the future of OpenAI, Brockman wrote about wanting to become a billionaire: “Financially what will take me to $1B?”  “Why didn’t you take the $29 billion and donate it to the

If you have spent time using AI coding agents — GitHub Copilot, Claude Code, Gemini CLI — you have probably run into this situation: you describe what you want, the agent generates a block of code that looks correct, compiles, and then subtly misses the actual intent. This “vibe-coding” approach can work for quick prototypes but becomes less reliable when building mission-critical applications or working with existing codebases. The issue, as GitHub frames it, is not the coding agent’s ability — it is the approach. Developers have been treating coding agents like search engines, when they should be treated more like literal-minded pair programmers who excel at pattern recognition but still need unambiguous instructions. To address this, GitHub has open sourced Spec-Kit — a toolkit designed to bring Spec-Driven Development (SDD) to AI coding workflows. The project currently has 90k+ stars and 8k+ forks on GitHub, and has become one of the faster-growing developer tooling repositories in recent memory. What is Spec-Driven Development? Spec-Driven Development inverts the traditional power structure of software development. Specifications do not serve code — code serves specifications. The Product Requirements Document (PRD) is not a guide for implementation; it is the source that generates implementation. In practice, this means you write a structured specification first — describing what you want to build and why, without specifying the tech stack — and then feed that into an AI coding agent as a grounding document. The spec becomes the source of truth that tools and AI agents use to generate, test, and validate code. The result is less guesswork, fewer surprises, and higher-quality code. This is distinct from “documentation-first” as traditionally practiced. SDD is not about writing exhaustive, dry requirements documents that nobody reads. It is not about waterfall planning or trying to predict the future through extensive planning exercises. And it is not about creating more bureaucracy that slows engineering teams down. The spec remains a living artifact — updated as requirements evolve, rather than filed away after project kickoff. What Spec-Kit Actually Includes Spec-Kithas two key components: the Specify CLI, a helper command-line tool that bootstraps projects for SDD by downloading official templates for the coding agent and platform of your choice; and a set of templates and helper scripts that establish the foundation for the SDD experience — defining what a spec looks like, what a technical plan encompasses, and how it all breaks down into individual tasks that an AI agent can execute. The CLI is written in Python and requires Python 3.11+. Installation via uv is the recommended method: Copy CodeCopiedUse a different Browser uv tool install specify-cli –from git+https://github.com/github/spec-kit.git@vX.Y.Z specify init <PROJECT_NAME> Once initialized, the agent has access to a set of slash commands that map directly to the SDD workflow. The core commands are: /speckit.constitution — establishes the project’s non-negotiable governing principles /speckit.specify — captures what you want to build, focused on the “what” and “why” without tech stack details /speckit.plan — generates the technical implementation plan given your chosen stack /speckit.tasks — breaks the plan into an actionable, dependency-ordered task list /speckit.taskstoissues — converts the generated task list into GitHub issues for tracking and execution /speckit.implement — executes those tasks using the AI coding agent There are also three optional commands for enhanced quality and validation: /speckit.clarify — surfaces underspecified areas through structured, sequential questioning before a technical plan is created (recommended before /speckit.plan to reduce rework downstream) /speckit.analyze — runs cross-artifact consistency and coverage analysis after /speckit.tasks and before /speckit.implement /speckit.checklist — generates custom quality checklists that validate requirements completeness, clarity, and consistency An important addition is constitution.md. In the SDD context, a constitution document establishes a set of non-negotiable principles for a project — testing conventions, CLI-first requirements, organizational design system standards. These are captured once and referenced throughout every subsequent development phase. GitHub Spec Kit How to Use GitHub Spec Kit: A Step-by-Step Guide Spec-Driven Development (SDD) with AI coding agents — from installing the CLI to running your first implementation. Follows the official workflow from the github/spec-kit repository. Step 1 of 10 Step 1 — Prerequisites Make sure you have the right tools installed Before installing the Specify CLI, you need four things on your machine. Spec Kit is cross-platform and works on Linux, macOS, and Windows. Python 3.11+ — download from python.org uv (recommended) or pipx for package management — install uv from docs.astral.sh/uv Git — download from git-scm.com A supported AI coding agent — Claude Code, GitHub Copilot, Gemini CLI, Cursor, Windsurf, Codex CLI, or any of the 29 supported integrations Why uv? It manages tool installations globally, keeps them in your PATH, and makes upgrading easy with uv tool list, uv tool upgrade, and uv tool uninstall. It’s the officially recommended method. 1 / 10 ← Back Next → Step 2 — Installation Install the Specify CLI from GitHub The only official Spec Kit package is published directly from the GitHub repository. Do not install from PyPI — any package there with the same name is not maintained by the Spec Kit team. # Persistent install (recommended) — replace vX.Y.Z with latest tag uv tool install specify-cli –from git+https://github.com/github/spec-kit.git@vX.Y.Z # Or using pipx pipx install git+https://github.com/github/spec-kit.git@vX.Y.Z # Verify installation specify version Check the Releases page for the latest tag (e.g. v0.8.4). Installing from main may include unreleased changes. One-time usage (no install): Run uvx –from git+https://github.com/github/spec-kit.git@vX.Y.Z specify init <PROJECT> to try without a persistent install. 2 / 10 ← Back Next → Step 3 — Initialize Bootstrap your project with specify init Navigate to your project folder and run specify init. The CLI detects which AI coding agent you have installed and sets up the right directory structure, templates, and commands automatically. # New project in a new folder specify init my-photo-app # Initialize inside an existing directory specify init . –integration claude # Skills mode for Claude Code and Codex CLI specify init . –integration codex –integration-options=”–skills” # Check all required tools are present specify check After this, your project

OpenAI released three new audio models through its Realtime API, each targeting a distinct capability in live voice applications: GPT-Realtime-2 for voice agents with reasoning, GPT-Realtime-Translate for live speech translation, and GPT-Realtime-Whisper for streaming transcription. Alongside the model releases, the Realtime API officially exits beta and is now generally available — a meaningful signal for developers who held off building production systems on it. All three models are available immediately through the OpenAI API and can be tested in the Playground. Together, they push voice applications past the basic question-and-answer loop — toward systems that can listen, reason, translate, transcribe, and act within a single conversation. GPT-Realtime-2: Voice Reasoning with a 128K Context Window The flagship release is GPT-Realtime-2, which OpenAI team describes as its first voice model with GPT-5-class reasoning. GPT-Realtime-2 can process harder requests, manage interruptions, and continue conversations naturally. OpenAI expanded the model’s context window from 32K to 128K tokens, allowing longer conversations and more complex tasks without losing context. Previous voice models frequently stalled on multi-step requests or dropped earlier context during longer sessions. GPT-Realtime-2 is specifically designed to keep the conversation moving while it reasons through a request. Developers can enable short preamble phrases — like “let me check that” or “one moment while I look into it” — so users know the agent is working on the request. The model can also call multiple tools at once and narrate what it’s doing while it does — so instead of dead air during a multi-step task, the user gets a running commentary. These features directly address one of the most common failure modes in deployed voice agents: awkward silence that makes the system feel broken. A particularly useful control for production builders is adjustable reasoning effort. Developers can dial reasoning intensity across five levels: minimal, low, medium, high, and xhigh. The default is “low” to keep latency down for simple requests, while tougher tasks can tap into more compute. This means teams can tune the performance-latency tradeoff at the session level depending on the use case — a quick customer lookup doesn’t need the same reasoning depth as a multi-step travel booking workflow. GPT-Realtime-2 also adds tone control. The model can adjust its speaking style depending on the situation — staying calm during problem-solving, shifting to empathetic when users are frustrated, and turning upbeat after a successful outcome. The model is also better at understanding industry-specific terminology, including healthcare vocabulary and proper nouns. On benchmarks, the gains are measurable. GPT-Realtime-2 with high reasoning scored 96.6% on Big Bench Audio, compared to 81.4% for GPT-Realtime-1.5 — a 15.2 percentage point improvement. GPT-Realtime-2 with xhigh reasoning scored 48.5% on Audio MultiChallenge instruction following, compared to 34.7% for GPT-Realtime-1.5. Big Bench Audio evaluates challenging reasoning capabilities in language models that support audio input. Audio MultiChallenge evaluates multi-turn conversational intelligence in spoken dialogue systems, including instruction following, context integration, self-consistency, and handling natural speech corrections. Pricing: GPT-Realtime-2 is priced at $32 per 1M audio input tokens ($0.40 for cached input tokens) and $64 per 1M audio output tokens. GPT-Realtime-Translate: Live Speech Translation Across 70+ Languages GPT-Realtime-Translate is a new live translation model that translates speech from 70+ input languages into 13 output languages while keeping pace with the speaker. Unlike GPT-Realtime-2, this model is a dedicated translation pipe — speech goes in one language and comes out in another. It is not a conversational agent; it is designed to convert one audio stream into another in real time. The distinction is important for developers choosing the right tool. If your application needs a bilingual customer support flow or a live interpreter for an in-person event, GPT-Realtime-Translate is the purpose-built option. If you need the model to also reason, call functions, or hold context across turns, GPT-Realtime-2 handles that. Pricing: GPT-Realtime-Translate is priced at $0.034 per minute. GPT-Realtime-Whisper: Streaming Transcription as People Speak GPT-Realtime-Whisper is a new streaming speech-to-text model built for low-latency speech-to-text — transcribing audio as people speak, so live products can feel faster, more responsive, and more natural. The original Whisper model was designed for completed chunks of audio, making it better suited for post-session transcription. GPT-Realtime-Whisper is the streaming counterpart, purpose-built for applications that need live output. For realtime transcription, gpt-realtime-whisper gives you controllable latency — lower delay settings produce earlier partial text, while higher delay settings can improve transcript quality. Use cases include live broadcast captions, meeting notes generated during the conversation, and voice agents that need to continuously understand the user rather than wait for turn-by-turn input. Pricing: GPT-Realtime-Whisper is priced at $0.017 per minute. Architecture Patterns and New Voices Developers can choose between three session types depending on the use case: a voice-agent session when the application needs an assistant that responds to the user, a translation session when the application needs an interpreter, and a transcription session when text from audio is needed without model-generated responses. On the voice output side, two new voices, Cedar and Marin, join the API roster exclusively with this release. All three models — GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper — are available now through the OpenAI Realtime API, which is generally available starting today. Key Takeaways GPT-Realtime-2 brings GPT-5-class reasoning to voice with a 128K context window, five-level adjustable reasoning effort, tone control, parallel tool calls, and interruption recovery On Big Bench Audio, GPT-Realtime-2 (high) scores 96.6% vs. 81.4% for GPT-Realtime-1.5; on Audio MultiChallenge, the xhigh variant scores 48.5% vs. 34.7%. GPT-Realtime-Translate handles live speech translation across 70+ input languages into 13 output languages at $0.034/min GPT-Realtime-Whisper streams transcription in real time with controllable latency at $0.017/min The Realtime API exits beta and goes generally available today alongside two new voices, Cedar and Marin Check out the Full Technical Details here. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect

arXiv:2605.06241v1 Announce Type: new Abstract: Reinforcement learning has become the standard for improving reasoning in large language models, yet evidence increasingly suggests that RL does not teach new strategies; it redistributes probability mass over solutions the base model already contains. In this work, we ask: if RL merely steers the model toward paths it already knows, is the RL optimization loop itself necessary? Through token-level analysis across multiple model families and RL algorithms, we find that RL’s beneficial footprint is a sparse, predictable correction concentrated at high-entropy decision points where the model is uncertain which branch to take. Only 1–3% of token positions are affected, the promoted token always lies within the base model’s top-5 alternatives, and targeted corrections at those few positions causally recover a large fraction of RL’s accuracy gain, while random corrections fail. The base model’s own entropy identifies these positions without any RL-trained model, and the entire correction is low-dimensional, representable in a tiny fraction of model parameters. These findings reframe reasoning improvement as sparse policy selection, not capability acquisition. We translate this insight into ReasonMaxxer, a minimal RL-free method that applies contrastive loss only at entropy-gated decision points, using a few hundred base-model rollouts and no online generation. Across three model families, six scales, and six math reasoning benchmarks, ReasonMaxxer matches or exceeds full RL performance while requiring only tens of problems and minutes of single-GPU training, a reduction in training cost of roughly three orders of magnitude.

arXiv:2605.05818v1 Announce Type: cross Abstract: Retrieval-Augmented Generation (RAG) enables large language models (LLMs) to leverage external knowledge, but also exposes valuable RAG databases to leakage attacks. As RAG systems grow more complex and LLMs exhibit stronger instruction-following capabilities, existing studies fall short of systematically assessing RAG leakage risks. We present LeakDojo, a configurable framework for controlled evaluation of RAG leakage. Using LeakDojo, we benchmark six existing attacks across fourteen LLMs, four datasets, and diverse RAG systems. Our study reveals that (1) query generation and adversarial instructions contribute independently to leakage, with overall leakage well approximated by their product; (2) stronger instruction-following capability correlates with higher leakage risk; and (3) improvements in RAG faithfulness can introduce increased leakage risk. These findings provide actionable insights for understanding and mitigating RAG leakage in practice. Our codebase is available at https://github.com/yeasen-z/LeakDojo.

Technology is changing the way we make babies. The pioneering work of the scientists who invented IVF led to the birth of the first “test tube baby” in 1978. We’ve come a long, long way since then. This week, I’ve been working on a piece about the cutting edge of IVF technologies and what’s coming next. Think AI and robots and, potentially, gene-edited embryos. My reporting has also made me think about just how much progress has been made in the last five decades. Clinicians have improved hormonal treatments. Embryologists have devised ways to culture embryos in the lab for longer. IVF clinics today offer multiple genetic tests for embryos. In recent years, we’ve had reports of babies born with DNA from three people, babies born following “IVF on wheels,” babies born from decades-old embryos, and even babies “conceived” with the aid of a sperm-injecting robot. The technology has also had a huge social impact. It has allowed for changes in the structure of families and provided more reproductive choices for would-be parents. So this week, let’s consider the technologies that have transformed babymaking. Alan Penzias, a reproductive endocrinologist at Boston IVF, has been working in IVF since the early 1990s. In those days, his lab at Yale would collect a person’s eggs, fertilize them, and culture any resulting embryos for two days, until the embryos had two or four cells. The embryos couldn’t survive any longer outside a body, so they’d be transferred to the uterus at that point. All of them. Even if there were, say, five embryos in total. Typical healthy patients could expect a live birth rate of 12% to 15%, he says. Then Penzias heard that other teams were managing to culture embryos for three days. “We thought, No, that’s not possible,” he recalls. He learned that scientists had achieved this by tinkering with the culture medium—the nutrient-rich fluid the embryos are grown in. Those three-day embryos, which had around six to 10 cells, seemed to have a better chance of resulting in a live birth. The teams culturing embryos for longer saw their success rates climb to 25% among similar patient groups, says Penzias. Again, he couldn’t believe it. “We thought they were making it up,” he says. In the years since, teams have made more improvements to culture medium. Today, most IVF embryos are cultured for five or six days—a point at which they have 80 to 100 cells. The culturing process can act a little like a stress test—the embryos that make it to day six are generally more likely to go all the way and develop into a healthy baby. Over the same period, advances in other technologies have opened up the options for what we can do with those embryos. Scientists learned they were able to freeze embryos and use them at a later date. A little over a decade ago, clinics shifted to a “vitrification” approach that rapidly cools the embryos to a glassy state. Vitrified embryos are more likely to survive freezing and thawing, so this approach quickly caught on. As a result, doctors no longer needed to transfer multiple embryos at once. This made it less likely that patients would have twins or triplets, which can increase the risk of pregnancy complications. Vitrification has also made IVF safer in other ways, including by affording patients a bit of time between fertility treatments. The hormonal treatments used in the first phase of IVF are designed to increase the production of mature eggs that can be collected. These treatments carry a small risk of a condition called ovarian hyperstimulation syndrome (OHSS), which in rare cases can be life-threatening. The ability to freeze all your embryos and use them at a later date is thought to give the body a chance to recover from hormonal treatment and reduces the risk of OHSS. And because clinics are now able to culture embryos for up to a week, they can take a few of the 100 or so cells and send them for genetic testing before freezing the embryos. People undergoing IVF can get genetic readouts of all the embryos before deciding which to implant. (It is worth noting, however, that these testing technologies are not perfect.) “Those are really radical changes, and we take them for granted,” says Penzias. These technologies have also changed the function of IVF. What was once a treatment for infertility is now used to preserve fertility. People who want to delay parenthood can opt to freeze their eggs or embryos and use them later. They might opt to transfer one embryo in a year’s time and a second several years later. “We’ve been able to empower women to be able to have much more reproductive choice and get more reproductive mileage from a single IVF cycle,” says Penzias. People who are about to undergo cancer treatments that might damage the testes or ovaries can opt to store their eggs or sperm ahead of time, too. Scientists have even been able to preserve pieces of ovarian and testicular tissue and reimplant them later, enabling recipients to have healthy babies. Today, more people than ever have access to safe IVF options that offer multiple paths to parenthood. Those options look set to expand. But if you want to find out more about the AI and IVF robots, you’ll have to read this week’s story, here! This article first appeared in The Checkup, MIT Technology Review’s weekly biotech newsletter. To receive it in your inbox every Thursday, and read articles like this first, sign up here.