Author: Naveed Ahmad

AI dictation apps have come a great distance in a short while. For years they had been sluggish and inaccurate — except you spoke with a specific accent and enunciated clearly. Advances in massive language fashions (LLMs) and speech-to-text fashions have modified that, producing techniques that may decipher speech extra precisely whereas retaining sufficient context to format the textual content appropriately. Builders have additionally in-built options to mechanically take away filler phrases, repair stumbles, and deal with punctuation — outputting textual content that wants far fewer edits. With dozens of such apps now in the marketplace, we’ve rounded up our…

Europe ought to be recognized for BottleCap AI, not bottle cap memes. With its tongue-in-cheek title, this Prague-based AI startup is likely one of the groups that VCs suppose you need to know. It isn’t that European startups by no means minimize by way of the noise — Lovable and Mistral AI are proof of it. However there are a lot of extra that don’t have 9 digits in annual recurring income but and that insiders are nonetheless monitoring very intently. That’s the place this record is available in. Over the previous few weeks, we requested traders at a few…

A gunman tried to enter the White Home Correspondents’ Dinner in Washington, DC, final weekend, whereas President Donald Trump, Vice President JD Vance, and different administration officers have been in attendance. Media reviews and Trump himself shortly recognized the suspected shooter as 31-year-old engineer and pc scientist Cole Tomas Allen. The California resident was arrested on the scene on Saturday and appeared Monday within the US District Courtroom for the District of Columbia to face three federal costs: trying to assassinate the president, transportation of a firearm in interstate commerce, and discharge of a firearm throughout a criminal offense of…

On this tutorial, we discover the lambda/hermes-agent-reasoning-traces dataset to grasp how agent-based fashions assume, use instruments, and generate responses throughout multi-turn conversations. We begin by loading and inspecting the dataset, inspecting its construction, classes, and conversational format to get a transparent thought of the obtainable info. We then construct easy parsers to extract key elements reminiscent of reasoning traces, device calls, and power responses, permitting us to separate inside considering from exterior actions. Additionally, we analyze patterns reminiscent of device utilization frequency, dialog size, and error charges to higher perceive agent conduct. We additionally create visualizations to focus on these…

Uber has a long-term ambition that goes effectively past shuttling passengers: the corporate finally desires to outfit its human drivers’ automobiles with sensors to absorb real-world knowledge for autonomous car (AV) corporations — and probably different corporations coaching AI fashions on physical-world eventualities. Praveen Neppalli Naga, Uber’s chief expertise officer, revealed the plan in an interview at TechCrunch’s StrictlyVC event in San Francisco on Thursday evening, describing it as a pure extension of a nascent program the corporate introduced in late January referred to as AV Labs. “That’s the path we need to go finally,” Naga stated of equipping human…

When you’ve got been working reinforcement studying (RL) post-training on a language mannequin for math reasoning, code technology, or any verifiable activity, you could have virtually definitely stared at a progress bar whereas your GPU cluster burns by way of rollout technology. A team of researchers from NVIDIA proposes a precise fix by integrating speculative decoding into the RL coaching loop itself, and do it in a manner that preserves the goal mannequin’s actual output distribution. The analysis workforce built-in speculative decoding straight into NeMo RL v0.6.0 with a vLLM backend, delivering lossless rollout acceleration at each 8B and projected…

Meta has acquired humanoid robotics startup Assured Robot Intelligence (ARI) for an undisclosed sum, the social media large stated. “We acquired Assured Robotic Intelligence, an organization on the frontier of robotic intelligence designed to allow robots to grasp, predict, and adapt to human behaviors in advanced and dynamic environments,” a Meta spokesperson instructed TechCrunch in an emailed assertion. ARI’s crew, together with its co-founders, will be part of Meta’s AI unit, the Superintelligence Labs analysis division. ARI had raised an undisclosed seed spherical from AI seed agency AIX Ventures. The startup was constructing basis fashions for humanoid robots to carry…

The bottleneck in constructing higher AI fashions has by no means been compute alone — it has at all times been information high quality. Meta AI’s RAM (Reasoning, Alignment, and Reminiscence) group is now addressing that bottleneck instantly. Meta researchers have launched Autodata, a framework that deploys AI brokers within the position of an autonomous information scientist, tasked with iteratively constructing, evaluating, and refining coaching and analysis datasets — with out counting on expensive human annotation at each step. And the outcomes, examined on advanced scientific reasoning issues, present that this method doesn’t simply match classical artificial information technology strategies…

Amjad Masad has been constructing Replit for a decade, however the final 18 months have been one thing else completely. The AI coding assistant firm went from $2.8 million in income in all of 2024 to monitoring towards what Masad describes as a billion-dollar annual run fee. At TechCrunch’s sold-out StrictlyVC occasion in San Francisco on Thursday night time, we coated a number of floor in a short while, starting with the query everybody within the business is asking proper now: in a world the place rival Cursor is reportedly in talks to be acquired by SpaceX for $60 billion,…

EPOCHS = 15 decide = torch.optim.AdamW(mannequin.parameters(), lr=1e-3, weight_decay=1e-4) sched = torch.optim.lr_scheduler.CosineAnnealingLR(decide, T_max=EPOCHS) loss_fn = nn.MSELoss() hist = {“tr”: [], “va”: [], “r”: []} def pearson(a, b): a, b = a – a.imply(), b – b.imply() return (a*b).sum() / (a.norm()*b.norm() + 1e-8) print(“n” + “=”*64) print(f”{‘Epoch’:>5} | {‘prepare’:>9} | {‘val’:>9} | {‘val_r’:>7}”) print(“=”*64) for ep in vary(EPOCHS): mannequin.prepare(); tr = [] for batch in train_loader: x, y = prep(batch) loss = loss_fn(mannequin(x), y) decide.zero_grad(); loss.backward() torch.nn.utils.clip_grad_norm_(mannequin.parameters(), 1.0) decide.step(); tr.append(loss.merchandise()) sched.step() mannequin.eval(); va, P, T = [], [], [] with torch.no_grad(): for batch in val_loader: x, y = prep(batch); p = mannequin(x)…