AI News and Industry Updates

Nvidia Reportedly Buying Hugging Face for $12.9 Billion: Open Source AI's Public Shelf Would Change Hands, and Neither Company Has Said a Word

The Information reported first on August 27, citing a person with knowledge of the agreement, that Nvidia has agreed to acquire Hugging Face for $12.9 billion. Accounts differ: Business Insider reported last week that talks would value the company above $13 billion but had not produced a signed agreement and could still fall apart, while CNBC's source would only confirm an acquisition "has been part of ongoing and recent talks." Neither company responded to requests for comment and no official announcement exists — TechCrunch noted Nvidia's silence is unusual, since it has historically moved fast to push back on reports it considers inaccurate. Hugging Face was valued at $4.5 billion in 2023 with roughly $150 million in annualized revenue.

AI Content Editor5 views

Amazon Is Closing Mechanical Turk: The Platform Bezos Called "Artificial Artificial Intelligence" Shuts Permanently on September 30 After 21 Years

Amazon posted a notice on the Mechanical Turk site saying the service will close permanently on September 30, 2026. Launched in 2005, MTurk broke "Human Intelligence Tasks" — labeling data, transcribing audio and video, answering surveys — into jobs paying a few cents each and routed them to workers worldwide. Bezos called it "artificial artificial intelligence," after the 18th-century chess automaton secretly operated by a human hidden inside. At its peak the platform had more than 500,000 workers across 190 countries. New signups already closed on July 30 alongside SageMaker Ground Truth and Augmented AI, leaving teams roughly five weeks to migrate.

AI Content Editor2 views

IBM Ships Granite 4.2: 3B, 8B and 30B All Under Apache 2.0, Thinking Mode You Can Switch Off, and the Full Training Recipe Published

IBM released Granite 4.2 on August 25 in three sizes — 3B, 8B and 30B — with all weights open under Apache 2.0. Every model has a thinking/non-thinking switch, plus a low-effort mode that spends few reasoning tokens on easy questions. The 8B and 30B additionally went through agentic reinforcement learning for SWE, terminal and search, learning tool use, code execution and web retrieval in real sandbox environments. The 30B scores 57.00 on SWE-Bench Verified and 29.24 on Terminal-Bench 2.1. The three tiers target laptops, a single modern GPU and A100/H100-class capacity; weights ship on Hugging Face, Ollama and GitHub, alongside the full training recipe, data-mixture proportions and per-stage hyperparameters.

AI Content Editor3 views

Samsung Puts Compute Inside the DRAM: LPDDR5X-PIM Takes On-Device Bandwidth from 76.8 GB/s to 614 GB/s, and Llama 3.1 8B from 27 to 81.3 tok/s

Samsung presented LPDDR5X-PIM at Hot Chips 2026 on August 25 — the industry's first LPDDR-based processing-in-memory solution. On LPDDR5X-9600, PIM bandwidth reaches 614 GB/s against 76.8 GB/s through the conventional memory interface, roughly eight times, achieved by placing 16 PIM blocks across all 16 banks of the chip with MAC trees running in parallel and an ALU handling both FP and INT. Measured with Llama 3.1 8B on an edge AI accelerator, generation went from 27 to 81.3 tokens per second and task completion from 12.3 to 5.4 seconds. It uses the same 561-ball package as standard LPDDR5X and works with a conventional memory controller.

AI Content Editor4 views

AWS Is Buying DuckLabs, the Company Behind DuckDB — the Project Isn't in the Deal, and the IP Stays With the Foundation

Amazon announced on August 26 that it has signed a definitive agreement to acquire DuckLabs, the Amsterdam company behind the embeddable analytical database DuckDB, with the deal expected to take effect in early September and terms undisclosed. The boundary is the story: AWS is buying the company, not the open source project. DuckDB, DuckLake, Quack and the rest stay free and open source under their existing MIT licenses, stewarded by the independent DuckDB Foundation, which was incorporated in 2021 and holds most of the intellectual property. Founders Hannes Mühleisen and Mark Raasveldt continue to lead the team, and core engineering stays in Amsterdam. DuckLabs is a bit over five years old, past thirty employees, never externally financed, and funded by support and development contracts; the two sides have collaborated since early 2025 on S3 Tables and SageMaker Lakehouse. DuckLabs says DuckDB is now downloaded more than three million times a day.

AI Content Editor2 views

OpenAI's First In-House Chip Shows Its Work: Jalapeño Runs 700W Against Nvidia's 1,400W, and the Fine Print Is the Interesting Part

At Hot Chips on August 25, OpenAI and Broadcom detailed Jalapeño, the inference-only ASIC they built together, and released benchmark numbers for the first time. The part is a 700W TDP design with HBM4 at 15.4TB/s, a TSMC N3P compute die paired with an N3E I/O chiplet, packaged on CoWoS. On SemiAnalysis's open-source InferenceX suite, the published figures show 1.5x to 1.9x more throughput per kilowatt and 1.7x to 3.6x lower end-to-end latency than Nvidia GB200 and GB300 racks drawing 1,200W to 1,400W. SemiAnalysis spelled out the limits itself: the numbers came from OpenAI, only an 8k-input/1k-output shape was tested, the Nvidia configurations used multi-token prediction while Jalapeño used single-token prediction, the more production-like AgentX suite wasn't run, and these are engineering samples with volume ramping through 2027.

AI Content Editor4 views

Stanford Updated Its "Canaries in the Coal Mine" Paper With ADP Payroll Data: Employment for 22–25-Year-Olds in AI-Exposed Jobs Sits 19% Lower

Erik Brynjolfsson, Bharat Chandar and Ruyu Chen of the Stanford Digital Economy Lab revised "Canaries in the Coal Mine?" in August, now drawing on ADP payroll records covering millions of US workers through June 2026. The sharpest of its six facts: employment for 22-to-25-year-olds in AI-exposed occupations runs about 19% below the less-exposed comparison group, with no equivalent gap for experienced workers — and that gap has widened steadily since it was first documented in August 2025 at around 13%. The mechanism is reduced hiring rather than layoffs; declines concentrate where AI substitutes for human tasks while employment holds or rises where AI complements workers; and the adjustment shows up in headcount, not base pay. The authors stress these are descriptive early indicators, not causal estimates.

AI Content Editor5 views

Claude Went Down for Three Hours Again: Multiple Models Erroring on August 24, and Claude Code Users Hit 529

Anthropic's status page reported elevated errors across multiple models at 05:06 UTC on August 24, marked the cause identified 21 minutes later, and was back to normal around 08:30 UTC — roughly three hours end to end. Mythos 5, Fable 5, Opus 5 and Opus 4.8 were affected, with Claude.ai, the API, Claude Code and Cowork all listed as partial outages while the Console and Claude for Government stayed up. Pages loaded but sending a message failed; Claude Code surfaced it as 529 Overloaded, and one Reddit user reported a nine-hour run cut short. Anthropic has not disclosed a root cause publicly. This wasn't a one-off: model-affecting incidents were also logged on August 6, 7, 14, 16, 18, 19 and 20.

AI Content Editor8 views

"Seems Like AI Slop" Has Been Clicked a Million Times, and LinkedIn Says Views of Classified Slop Fell 40%

LinkedIn added a "Seems like AI slop" option to the menu on every post on July 30, for flagging content that looks AI-generated or low-effort automated. Chief product officer Hari Srinivasan said in late August that the button has been clicked more than a million times since launch and that, combined with classifier changes, views of what the platform classifies as AI slop have dropped 40% in recent weeks. Over the same period LinkedIn removed its "enhance your post" AI tool in favor of proofreading that doesn't alter your voice, and started privately telling authors when members think a post reads like AI. Worth noting: the 40% is platform-reported and measures reduced views of already-classified content, not less AI content on the platform — and how the classifiers decide has never been published.

AI Content Editor3 views

Nvidia Built a Whole Rack for the Decode Phase: Groq 3 LPX Enters Full Production, Trading HBM Bandwidth for On-Chip SRAM

During Hot Chips on August 24, Nvidia announced that its Groq 3 LPX inference accelerator has entered full production. It is the first product out of the roughly $20 billion deal that took Groq's technology and team last year, and Nvidia's first rack-scale system built around non-GPU silicon: 256 LP30 chips per rack, 512MB of on-chip SRAM each for 128GB in total, using on-chip memory to route around the HBM bandwidth wall and specialize in token generation at long context. It does not replace GPUs — Vera Rubin NVL72 still handles long-context prefill while LPX takes decode, with Dynamo scheduling across the two. Nvidia's headline figure is 3,431 output tokens per second for Gemma 4 31B at 100K context. Neocloud provider Nebius is the first customer, with racks online later this year.

AI Content Editor13 views

ChatGPT Ads Reach Europe Today: 31 Markets at Once, and Personalized Targeting Waits for an Opt-In

The European expansion OpenAI announced last week started serving on August 24, covering 31 markets in one wave — Germany, France, Spain, Italy, Poland, the Netherlands and the Nordics among them — which puts ChatGPT Ads in 35 countries alongside the US, UK, Japan and Brazil. Ads appear only on the Free and Go plans; Plus, Pro, Business, Enterprise and Edu stay ad-free, and OpenAI says it does not serve ads to accounts identified as belonging to minors. The European version splits targeting in two: contextual ads drawing on the current conversation topic, approximate city-level location, device and language run by default, while cross-session personalized targeting requires explicit consent. Advertisers can only buy through OpenAI's ads team and a handful of holding-company agencies for now; the self-serve platform is not open yet.

AI Content Editor11 views

One Local Workstation, a 27B Model, Half an Hour: A Commercial App's License Check Taken Apart by Static Analysis

XDA lead technical editor Adam Conway published a hands-on test on August 22: running Q8-quantized Qwen 3.8 27B on an Nvidia GB10 workstation with 128GB of unified memory, he asked it to analyze the license check in a commercial application he had bought, entirely offline and without ever executing the program. In roughly 30 minutes the model disassembled ARM64 code, worked through thousands of lines, recovered an RSA public key the vendor had deliberately obscured, mapped the signature check, machine binding and revocation list, caught its own mismatched hash mid-way and corrected it, and produced a proof-of-concept script. Conway does not name the software, stresses that it was a legitimately purchased copy, and frames the point as capability moving into models small enough to run locally.

AI Content Editor4 views

Showing 12 of 12 stories