Hi there, this is your daily ☕️ Techpresso.
In today's Techpresso:
👀 Amazon launches dashboard to monitor employee productivity
🛰️ SpaceX approved to launch 7500 more Starlink satellites
🤖 Insiders say DeepSeek V4 will beat Claude and ChatGPT
🔒 Anthropic blocks unauthorized third-party access to Claude
📡 Cloudflare defies Italy’s Piracy Shield
📚 Study proves AI models can reproduce entire copyrighted books
🎁 + 9 other news you might like
🧰 + 6 trending tools
📚 + 6 trending papers
👀 Amazon launches dashboard to monitor employee productivity LINK
Amazon has launched a new dashboard that lets managers track how often employees come into the office and how long they stay, using badge tap data collected over eight weeks.
The system groups workers into categories like "low-time" and "zero" badgers, and it flags people who frequently enter buildings they aren't assigned to, though managers decide what action to take.
The stricter tracking follows Amazon's five-day return-to-office mandate from 2024, when an internal survey found 73% of employees were considering quitting over the policy change.
🛰️ SpaceX approved to launch 7500 more Starlink satellites LINK
The Federal Communications Commission approved SpaceX to launch 7,500 more second-generation Starlink satellites, which doubles the company's total Gen2 authorization to 15,000 satellites for its internet network.
The FCC had delayed this approval since December 2022 to address concerns about orbital debris and space safety, and the new order allows SpaceX to operate across multiple frequency bands.
SpaceX also plans to lower about 4,400 existing satellites from 550 kilometers to 480 kilometers during 2026, saying the lower altitudes have less debris that satellites could hit.
🤖 Insiders say DeepSeek V4 will beat Claude and ChatGPT LINK
Sources say DeepSeek plans to release its V4 model around mid-February, and internal tests reportedly show it outperforms both Claude and ChatGPT, especially when handling extremely long code prompts.
V4 is a hybrid model focused on coding tasks for enterprise developers, and to claim dominance it would need to beat Claude Opus 4.5, which holds the SWE-bench Verified record at 80.9%.
DeepSeek's potential edge comes from a new training method called Manifold-Constrained Hyper-Connections, which lets models scale larger without becoming unstable, even with limited access to advanced chips due to U.S. export restrictions.
🔒 Anthropic blocks unauthorized third-party access to Claude LINK
Anthropic has added new technical safeguards to stop third-party apps from pretending to be its official Claude Code tool, which let users access Claude AI models with better pricing and higher limits.
The crackdown broke workflows for users of OpenCode and similar tools that spoofed Claude Code headers to run automated coding tasks on flat-rate consumer subscriptions instead of paying per-token API prices.
Separately, Anthropic cut off access for staff at Elon Musk's xAI lab, who had been using Claude models through the Cursor IDE, citing terms of service rules against competitors training rival AI systems.
📡 Cloudflare defies Italy’s Piracy Shield LINK
Cloudflare has been fined 14.2 million euros by Italy's communications agency AGCOM for refusing to block access to pirate sites through its 1.1.1.1 DNS service under the Piracy Shield law.
Cloudflare argued that filtering roughly 200 billion daily DNS requests would increase latency and hurt legitimate sites, but AGCOM rejected this, claiming the targeted IP addresses were only used for copyright infringement.
Researchers found hundreds of legitimate websites were affected by Piracy Shield blocking, while illegal streamers continued to evade enforcement by exploiting the abundance of available address space online.
📚 Study proves AI models can reproduce entire copyrighted books LINK
A new study from Stanford and Yale researchers shows that major AI models can reproduce nearly complete copies of copyrighted books like Harry Potter and the Sorcerer's Stone when prompted with specific techniques.
Claude 3.7 Sonnet reproduced 95.8% of the Harry Potter book after jailbreak prompts, while Gemini 2.5 Pro and Grok 3 output 76.8% and 70.3% respectively without any jailbreaking needed at all.
The findings directly challenge AI companies' legal defense that models do not store training data, which could affect ongoing copyright lawsuits and potentially expose the industry to billions in damages.
Other news you might like
- OpenAI and SoftBank announce $1 billion investment in SB Energy as part of massive AI buildoutLINK
- North Korean hackers using malicious QR codes in spear phishing, FBI warnsLINK
- Google: Don’t make “bite-sized” content for LLMs if you care about search rankLINK
- “Ungentrified” Craigslist may be the last real place on the InternetLINK
- AI benchmarks are broken and the industry keeps using them anyway, study findsLINK
- Burry explains why Oracle is on his short listLINK
- Google moonshot spinout SandboxAQ claims an ex-exec is attempting ‘extortion’LINK
🧰 Trending tools
Lend Me Your Ears: a memory game that trains users to recognize and play musical intervals and melodies on piano without sheet music.LINK
Repo Prompt: generates context-aware prompts from your codebase structure to help AI coding agents understand your project and produce more accurate results.LINK
Settle It: a no-login polling tool that creates shareable decision polls via a single link, eliminating endless group chat debates about simple choices.LINK
Boom for Mac: a video presentation tool that lets you record polished tutorials or present live with pro layouts and transitions, eliminating post-production editing entirely.LINK
Burner Note: a zero-knowledge encrypted messaging tool that auto-deletes notes after one read, ensuring truly private communication without server-side data retention.LINK
Scroll Wikipedia like TikTok: generates Wikipedia-style content in a vertical scrolling feed using LLMs, with HTML rendered just-in-time into an infinite social media interface.LINK
📚 Trending papers & reports
Language models leak training data verbatim: researchers extracted entire copyrighted books from ChatGPT by repeatedly prompting it to continue generating text sequences.LINK
Electronic chip solves optimization problems using magnetic physics: researchers built hardware that finds best solutions 10,000 times faster than software by mimicking how magnets align.LINK
LLM agents stay compromised after one poisoned memory: researchers found that AI assistants can be permanently hijacked when attackers inject a single malicious memory into their experience database.LINK
Single pixel changes fool image recognition: researchers showed that changing just one pixel in an image can trick AI into completely misidentifying objects with high confidence.LINK
Living cells turned into computers: researchers engineered bacteria to process information like brain cells, creating biological circuits that could sense and respond to environments.LINK
AI chatbot reduces online hostility when correcting false claims: researchers found Grok-generated replies corrected misinformation while reducing personal attacks by 60% compared to typical user responses.LINK
Want to get the latest news differently? Find us on:
See you tomorrow for a new dose of ☕️ Techpresso!
.png){kind=small}
.png){kind=small}
.png){kind=small}