Cohere

Cohere AI

Cohere AI

Cohere AI

Cohere AI

Cohere AI

Cohere AI

Cohere AI

Cohere AI

Cohere AI

Cohere AI

Cohere Transcribe: Speech Recognition

Cohere's open-weight ASR model hits 5.4% word error rate — low enough to replace speech APIs in production pipelines

Enterprises building voice-enabled workflows have had limited options for production-grade transcription: closed APIs with data residency risks, or open models that trade accuracy for deployability. Cohere's new open-weight ASR model, Transcribe, is built to compete on all four key differentiators — contextual accuracy, latency, control and cost. Cohere says that Transcribe outperforms current leaders on accuracy — and unlike closed APIs, it can run on an organization's own infrastructure. Cohere, which can be accessed via an API or in Cohere’s Model Vault as cohere-transcribe-03-2026, has 2 billion parameters and is licensed under Apache-2.0. The company said Transcribe has an average word error rate (WER) of just 5.42%, so it makes fewer mistakes than similar models. It’s trained on 14 languages: English, French, German, Italian, Spanish, Greek, Dutch, Polish, Portuguese, Chinese, Japanese, Korean, Vietnamese and Arabic. The company did not specify which Chinese dialect the model was trained on. Cohere said it trained the model “with a deliberate focus on minimizing WER, while keeping production readiness top-of-mind.” According to Cohere, the result is a model that enterprises can plug directly into voice-powered automations, transcription pipelines, and audio search workflows. Self-hosted transcription for production pipelines Until recently, enterprise transcription has been a trade-off — closed APIs offered accuracy but locked in data; open models offered control but lagged on performance. Unlike Whisper, which launched as a research model under MIT license, Transcribe is available for commercial use from release and can run on an organization's own local GPU infrastructure. Early users flagged the commercial-ready open-weight approach as meaningful for enterprise deployments. Organizations can bring Transcribe to their own local instances, since Cohere said the model has a more manageable inference footprint for local GPUs. The company said they were able to

An incoherent Rust

LangChain's CEO argues that better models alone won't get your AI agent to production

As models get smarter and more capable, the "harnesses" around them must also evolve. This "harness engineering" is an extension of context engineering, says LangChain co-founder and CEO Harrison Chase in a new VentureBeat Beyond the Pilot podcast episode. Whereas traditional AI harnesses have tended to constrain models from running in loops and calling tools, harnesses specifically built for AI agents allow them to interact more independently and effectively perform long-running tasks. Chase also weighed in on OpenAI's acquisition of OpenClaw, arguing that its viral success came down to a willingness to "let it rip" in ways that no major lab would — and questioning whether the acquisition actually gets OpenAI closer to a safe enterprise version of the product. “The trend in harnesses is to actually give the large language model (LLM) itself more control over context engineering, letting it decide what it sees and what it doesn't see,” Chase says. “Now, this idea of a long-running, more autonomous assistant is viable.” Tracking progress and maintaining coherence While the concept of allowing LLMs to run in a loop and call tools seems relatively simple, it’s difficult to pull off reliably, Chase noted. For a while, models were “below the threshold of usefulness” and simply couldn’t run in a loop, so devs used graphs and wrote chains to get around that. Chase pointed to AutoGPT — once the fastest-growing GitHub project ever — as a cautionary example: same architecture as today's top agents, but the models weren't good enough yet to run reliably in a loop, so it faded fast. But as LLMs keep improving, teams can construct environments where models can run in loops and plan over longer horizons, and they can continually improve these harnesses. Previously, “you couldn't really make improvements to the harness because you couldn't actually run the model in a harness,” Chase said. LangChain’s answer to this is Deep Agents, a customizable general-purpose harness. Built on L

I created a mathematical framework for AI Alignment and I would like to work with people in the alignment community as collaborators. I appreciate all the help and support I can get.

[Original Reddit post](https://www.reddit.com/r/ArtificialInteligence/comments/1rnfb3y/i_created_a_mathematical_framework_for_ai/) TRC: Trust Regulation and Containment A Predictive, Physics-Inspired Safety Framework for Large Language Models TRC: Trust Regulation and Containment A Predictive, Physics-Inspired Safety Framework for Large Language Models Kevin Couch Abstract Large language models exhibit structural failure modes—hallucination, semantic drift, sycophancy, and dyadic dissociation—that cause measurable harm, particularly to vulner- able users. TRC (Trust Regulation and Containment) is a two-layer, inference-time frame- work that combines a hard binary Trust Gate with a continuous, physics-inspired Ethical Rheostat operating directly on the model’s residual-stream activation vector. By tracking semantic momentum across layer depth and applying graduated, tensor-based geometric projections, TRC shifts safety enforcement from reactive post-generation filtering to a pre- dictive, self-correcting control law. The core is a stochastic differential equation—re-indexed to layer depth under an approx- imate Neural ODE interpretation—that augments the transformer’s natural forward flow with an ethical steering term derived from a compact set of contrastively extracted concept vectors. This revision introduces eight principal advances: (i) an adaptive gain law Λ+(l) whose gain response accelerates into danger and decelerates into safety without oscillation risk; (ii) a scalar Kalman filter with a clutch mechanism that closes the Bayesian momentum predictor implementation gap, provably optimal under the framework’s own Gaussian noise assumptions and decoupled from burst dynamics via federated regime handoff; (iii) a formal Itô stability condition giving implementers an analytical lower bound on λ0; (iv) replacement of the instantaneous jump operator with a continuous flow burst mechanism that preserves activation manifold geometry; (v) a calibration shunt reference Cref normalising all thresh- olds and gain coefficients against a known-safe baseline; (vi) a tempo efficiency framework unifying token cost, electrical cost, and coherence distortion into a single joint optimisa- tion objective; (vii) a signed gain architecture that partitions each concept projection into harmful and prosocial components, with detection and escalation operating exclusively on the harmful channel C+ to prevent adversarial prosocial suppression; and (viii) a Kalman clutch mechanism implementing federated estimation with deterministic Lyapunov stabil- ity during burst episodes and stochastic Lyapunov stability during nominal operation, with formally specified regime transitions. Stochastic perturbation is projected into the ethical subspace, making the Langevin diffusion interpretation exact rather than approximate. The framework is validated against chess dynamics, which constitute a well-studied discrete dy- namical system whose positional flow, tactical burst, and zugzwang properties map precisely onto TRC’s three-term master equation. Introduction Large language models exhibit a range of structural failure modes—hallucination, semantic drift, sycophancy, and dyadic dissociation—that can cause measurable harm, especially to vulnerable users. These phenomena arise not from reasoning errors but from the probabilistic nature of transformer sampling and the high-dimensional geometry of activation space. In this paper we present TRC (Trust Regulation and Containment), a two-layer, inference-time framework that blends hard decision gates with a continuous, physics-inspired correction engine operating directly on the model’s residual-stream activation vector. The central geometric insight motivating this revision is that the transformer’s residual stream traces a continuous path through a high-dimensional activation manifold. Safety failures are deformations of this manifold—crinkles in its geometry introduced by adversarial inputs, sycophantic drift, or escalating user distress. The correct response to a crinkle is not to teleport the activation to a safe location (which introduces new geometric incoherence) but to apply continuous corrective flow that works the deformation out smoothly, layer by layer, the way a craftsperson works aluminum foil back toward its intended shape. This insight drives the replacement of the previous instantaneous jump operator with the flow burst architecture and motivates the tempo efficiency framework that unifies all computational cost metrics under a single variable. This revision also introduces the Kalman clutch mechanism, which decouples the Bayesian momentum predictor from burst dynamics during high-gain corrective episodes. The system now operates as a federated estimation architecture with formally specified regime transitions: nominal tracking under stochastic Lyapunov stability, deterministic correction during burst episodes, and a principled re-engagement protocol with inflated covariance. Th

Trump’s War to Nowhere

The Israel–U.S. military campaign in Iran has killed more than 1,000 people since the assault began on February 28. A [war powers resolution](https://theintercept.com/2026/03/04/iran-war-powers-gottheimer-fetterman/) in the Senate to curb President Donald Trump’s ability to drag the U.S. into the war failed on Wednesday. Similarly, a measure in the House failed on Thursday. “This war is just a few days old and it’s escalating really quickly,” says Ali Gharib, senior editor at The Intercept. “It’s becoming a regional conflict,” as Iran retaliates and targets [U.S. bases](https://www.aljazeera.com/news/2026/3/5/drone-targets-us-base-in-iraq-as-iran-attacks-hit-region-amid-us-israel-war) as well as Israel and Gulf energy sites. This week on The Intercept Briefing, Gharib discusses the human and political toll of the Israel–U.S. war on Iran with co-host Jordan Uhl and journalist [Séamus Malekafzali](https://theintercept.com/staff/seamus-malekafzali/), who has been based in Paris and Beirut. [ Related ------- ### Sources Briefed on Iran War Say U.S. Has No Plans for What Comes Next](https://theintercept.com/2026/03/05/trump-iran-war-plan-cia/) “Trump has repeatedly failed to articulate anything even resembling coherent about why the U.S. got into this war,” says Gharib. He adds, “Marco Rubio even — who, again, not the sharpest tool in the shed, but usually has his shit pretty together — but in this case, he’s like changing his tune every two days because he has to keep up with Trump’s inanity about what the reasons for the war were.” The end game for Israel here, says Malekafzali, is they want “a state that is incapable of defending itself, a state that is no longer sovereign.” He adds, “If you are bombarding police stations, if you are bombarding hospitals and schools, border guards, when you are attacking the very fabric of any society as your main target, CENTCOM and the IDF together, that means that you are going toward state collapse.” “These are hard-won lessons over and over again for the United States — war after war, fallout, blowback. It just happens again and again. And yet we always seem to get leaders who are willing to run willy-nilly into these things,” says Gharib. Listen to the full conversation of The Intercept Briefing on [Apple Podcasts](https://podcasts.apple.com/us/podcast/the-intercept-briefing/id1195206601), [Spotify](https://open.spotify.com/show/2js8lwDRiK1TB4rUgiYb24), [YouTube](https://www.youtube.com/playlist?list=PLW0Gy9pTgVnvgbvfd63A9uVpks3-uwudj), or wherever you listen. **Transcript** -------------- **Jordan Uhl:** Welcome to the Interceptive Briefing, I’m Jordan Uhl. **Ali Gharib**: And I’m Ali Gharib. I’m a senior editor at The Intercept. **JU:** Today we’re going to talk about the growing war in the Middle East, specifically Iran. Last Saturday, Israel and the United States launched unprovoked attacks on Iran, and assassinated Supreme Leader [Ali Khamenei](https://www.theguardian.com/world/2026/mar/01/ayatollah-ali-khamenei-obituary) as well as several senior military officials. The Israel–U.S. strikes have continued on Iran, bringing the [death toll](https://www.aljazeera.com/news/2026/3/4/death-toll-in-iran-surpasses-1000-as-israel-us-strikes-continue) to more than 1,000 people since the assault began. On Thursday, the World Health Organization verified [13 attacks on health infrastructure](https://www.reuters.com/world/middle-east/who-says-has-it-has-verified-13-health-attacks-iran-2026-03-05/) that killed four health care workers. Ali, it feels like we’ve seen this playbook run before, but this time, it seems like they’re trying to distinguish what is and what isn’t a war. **AG**: This is like the sort of last readout of the idiot, when it comes to national security policy, is that you don’t need congressional approval. There’s no real stakes because this isn’t a war. This is part of a long history. It’s bipartisan. We’ve seen Democrats in office. We’ve seen Republicans in office. People are constantly starting these wars. They say they’re going to be limited strikes. Well, you know what? When you’re dropping bombs on another country and that country is attacking your military personnel in the area, that’s a textbook war. In the so-called [global war on terror](https://theintercept.com/collections/the-911-wars/), they could bullshit this and say, “Oh, we’re not going after armies. We’re going after these non-state actors and terrorist groups,” or whatever. But in this case, it’s like you’re literally attacking the leadership of another country and another country’s military. There’s just no way to bullshit this. This is war. It’s what it is. There’s [civilians dying](https://www.reuters.com/world/middle-east/how-many-people-have-been-killed-us-israel-war-iran-2026-03-03/). It’s th

Trump’s War to Nowhere

The Israel–U.S. military campaign in Iran has killed more than 1,000 people since the assault began on February 28. A [war powers resolution](https://theintercept.com/2026/03/04/iran-war-powers-gottheimer-fetterman/) in the Senate to curb President Donald Trump’s ability to drag the U.S. into the war failed on Wednesday. Similarly, a measure in the House failed on Thursday. “This war is just a few days old and it’s escalating really quickly,” says Ali Gharib, senior editor at The Intercept. “It’s becoming a regional conflict,” as Iran retaliates and targets [U.S. bases](https://www.aljazeera.com/news/2026/3/5/drone-targets-us-base-in-iraq-as-iran-attacks-hit-region-amid-us-israel-war) as well as Israel and Gulf energy sites. This week on The Intercept Briefing, Gharib discusses the human and political toll of the Israel–U.S. war on Iran with co-host Jordan Uhl and journalist [Séamus Malekafzali](https://theintercept.com/staff/seamus-malekafzali/), who has been based in Paris and Beirut. [ ## Related ### Sources Briefed on Iran War Say U.S. Has No Plans for What Comes Next](https://theintercept.com/2026/03/05/trump-iran-war-plan-cia/) “Trump has repeatedly failed to articulate anything even resembling coherent about why the U.S. got into this war,” says Gharib. He adds, “Marco Rubio even — who, again, not the sharpest tool in the shed, but usually has his shit pretty together — but in this case, he’s like changing his tune every two days because he has to keep up with Trump’s inanity about what the reasons for the war were.” The end game for Israel here, says Malekafzali, is they want “a state that is incapable of defending itself, a state that is no longer sovereign.” He adds, “If you are bombarding police stations, if you are bombarding hospitals and schools, border guards, when you are attacking the very fabric of any society as your main target, CENTCOM and the IDF together, that means that you are going toward state collapse.” “These are hard-won lessons over and over again for the United States — war after war, fallout, blowback. It just happens again and again. And yet we always seem to get leaders who are willing to run willy-nilly into these things,” says Gharib. Listen to the full conversation of The Intercept Briefing on [Apple Podcasts](https://podcasts.apple.com/us/podcast/the-intercept-briefing/id1195206601), [Spotify](https://open.spotify.com/show/2js8lwDRiK1TB4rUgiYb24), [YouTube](https://www.youtube.com/playlist?list=PLW0Gy9pTgVnvgbvfd63A9uVpks3-uwudj), or wherever you listen. ## **Transcript** **Jordan Uhl:** Welcome to the Interceptive Briefing, I’m Jordan Uhl. **Ali Gharib**: And I’m Ali Gharib. I’m a senior editor at The Intercept. **JU:** Today we’re going to talk about the growing war in the Middle East, specifically Iran. Last Saturday, Israel and the United States launched unprovoked attacks on Iran, and assassinated Supreme Leader [Ali Khamenei](https://www.theguardian.com/world/2026/mar/01/ayatollah-ali-khamenei-obituary) as well as several senior military officials. The Israel–U.S. strikes have continued on Iran, bringing the [death toll](https://www.aljazeera.com/news/2026/3/4/death-toll-in-iran-surpasses-1000-as-israel-us-strikes-continue) to more than 1,000 people since the assault began. On Thursday, the World Health Organization verified [13 attacks on health infrastructure](https://www.reuters.com/world/middle-east/who-says-has-it-has-verified-13-health-attacks-iran-2026-03-05/) that killed four health care workers. Ali, it feels like we’ve seen this playbook run before, but this time, it seems like they’re trying to distinguish what is and what isn’t a war. **AG**: This is like the sort of last readout of the idiot, when it comes to national security policy, is that you don’t need congressional approval. There’s no real stakes because this isn’t a war. This is part of a long history. It’s bipartisan. We’ve seen Democrats in office. We’ve seen Republicans in office. People are constantly starting these wars. They say they’re going to be limited strikes. Well, you know what? When you’re dropping bombs on another country and that country is attacking your military personnel in the area, that’s a textbook war. In the so-called [global war on terror](https://theintercept.com/collections/the-911-wars/), they could bullshit this and say, “Oh, we’re not going after armies. We’re going after these non-state actors and terrorist groups,” or whatever. But in this case, it’s like you’re literally attacking the leadership of another country and another country’s military. There’s just no way to bullshit this. This is war. It’s what it is. There’s [civilians dying](https://www.reuters.com/world/middle-east/how-many-people-have-been-killed-us-israel-war-iran-2026-03-03/). It’s the whole thing. It