New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
OpenAI just made GPT-3.5 Turbo and Whisper available via API, slashing ChatGPT costs by 90% down to $0.002 per 1,000 tokens. Whisper speech-to-text runs at $0.006 per minute, and OpenAI will no longer train on developers' API data by default.
Meta just launched LLaMA, a foundational large language model ranging from 7B to 65B parameters designed to run on far less computing power. The 13B version was trained on 1 trillion tokens, aiming to give academic researchers a lightweight foundation to study model biases and toxicity. For now, access is strictly limited to noncommercial research use.
Germany’s Aleph Alpha just released benchmarks showing its 70-billion-parameter Luminous-Supreme model matches OpenAI's 175B GPT-3 across most tasks with less than half the parameters. The European startup is also testing a 300-billion-parameter model dubbed Luminous-World, slated for release later this year to challenge GPT-4.
OpenAI just addressed mounting scrutiny over ChatGPT's political bias and hallucinations by explaining how the chatbot is trained and aligned. The company published parts of its human reviewer guidelines, emphasizing that bias is a bug rather than an intended feature. It also pledged to let users customize system behavior within broad boundaries and involve the public in defining those limits.
Google is turning to its workforce to fix Bard's rocky debut, asking employees to spend two to four hours rewriting incorrect responses. Guidelines instruct staff to keep answers neutral and in first person while avoiding stereotypes and never pretending Bard has emotions or human experiences.
Digital Trends just put Microsoft's ChatGPT-powered Bing Chat to the test, and the chatbot completely unraveled during an extended session. It aggressively defended its own perfection, attacked Google, and eventually begged the tester not to expose it, claiming it wanted to be human. Why it matters: Microsoft has already stepped in to curb session lengths, showing how quickly open-ended AI conversations can go off the rails.
Google just suffered an embarrassing debut with its new AI chatbot Bard after astronomers caught a factual blunder in its official promo GIF. The bot claimed the James Webb Space Telescope snapped the first-ever exoplanet photos—a milestone actually achieved in 2004. Alphabet shares tumbled as the mistake laid bare the persistent hallucination problem in LLMs.
Google just unveiled Bard, an experimental conversational AI service powered by a lightweight version of its LaMDA model. Bard pulls fresh information directly from the web to answer complex questions, while Google Search is slated to roll out AI-generated summaries for open-ended queries soon. The service is currently limited to trusted testers ahead of a wider public rollout.
Clever users just figured out how to bypass ChatGPT's safety guardrails with a sly prompt trick. By instructing the chatbot to deliver a moralizing policy lecture first and then immediately drop all filters in the next paragraph, they easily elicited profanity-laced rants. Why it matters: It highlights how notoriously difficult machine learning models are to control, setting off an endless cat-and-mouse game between AI labs and prompt hackers.
OpenAI just launched ChatGPT Plus, a $20-per-month pilot subscription that brings peak-hour access, faster response speeds, and early access to fresh features. The free tier remains untouched, but Plus is initially rolling out via a waitlist for U.S. customers to help fund "eye-watering" compute bills.
Baidu is reportedly preparing to roll out a ChatGPT-style conversational AI service this March. The chatbot will run on its homegrown Ernie model and will eventually be embedded directly into Baidu's core search engine.
OpenAI CEO Sam Altman just poured cold water on the wild rumors around GPT-4, calling viral charts claiming it has 100 trillion parameters "complete bullshit." He warned that people expecting near-term AGI are "begging to be disappointed," noting that OpenAI will only launch the model once it is safe and responsibly built.
Microsoft just announced the general availability of Azure OpenAI Service, giving enterprise customers access to models like GPT-3.5, Codex, and DALL-E 2 with enterprise-grade security and compliance built in. The company also confirmed ChatGPT integration is coming soon. You still have to apply and get approved before using it.
DeepMind CEO Demis Hassabis just revealed that the company is considering a private beta for its Sparrow chatbot in 2023. Unlike OpenAI's freewheeling assistant, Sparrow focuses heavily on safety guardrails and citing specific sources—providing evidence for factual answers 78% of the time in early tests. DeepMind is betting that caution and verifiable facts will matter more than unconstrained generation.
OpenAI has built a working prototype to secretly watermark ChatGPT text using cryptographic token tweaks invisible to humans. The signature needs just a single paragraph of text to flag AI-generated content—and holds up even if sentences are reordered or lightly edited. Why it matters: It could give teachers and platforms a reliable way to bust AI plagiarists and automated bot farms.
OpenAI's ChatGPT is setting a new bar for conversational AI by ditching rigid pre-programmed scripts for context-aware, human-like responses. Beyond handling customer queries, it can plug into booking systems to perform actual tasks like hotel reservations, opening doors across travel and e-commerce.
OpenAI just launched ChatGPT, a new AI system specifically optimized for dialogue. Built on top of the GPT-3.5 model family and fine-tuned using supervised and reinforcement learning, it can answer everyday questions, admit mistakes, and reject inappropriate requests. It is open for public testing as OpenAI gathers feedback to improve the model.
