OpenAI's new Lockdown Mode for ChatGPT disables web access, Deep Research, and Agent Mode to make data theft through prompt injection attacks harder. The mode doesn't fully prevent such attacks, it only blocks the final step in an exfiltration chain. Prompt injection remains an unsolved problem. The article ChatGPT's new Lockdown Mode lets you disable web access and more to protect sensitive data from prompt injection appeared first on The Decoder . ]]>
Bookkeepers do important work. But a bookkeeper is not the head of accounting. Marketers are responsible for anything the organization does that touches the market. But many people with ‘marketer’ in their title simply go to meetings and do tasks after the real work of marketing is already done. Some tech companies have hundreds of […]
Clive Chan, by his own account the second hardware employee in OpenAI's custom chip program, is moving to Anthropic. He brings experience from Tesla's Autopilot ASIC and the OpenAI-Broadcom partnership. The move comes as both companies prepare for their IPOs, and Anthropic is reportedly considering developing its own AI chips. The article Anthropic poaches OpenAI's second-ever chip engineer as both companies race toward IPOs appeared first on The Decoder . ]]>
Small language models fail at rare tasks because frequent ones constantly overwrite what they've learned. A new study with models ranging from 4 million to 4 billion parameters shows this mechanism in detail and offers a practical fix: instead of scaling up models, it may be enough to increase how often the target task appears in the training data. The article Researchers pinpoint why larger language models pick up skills that small ones miss appeared first on The Decoder . ]]>
<h1>New models</h1> <ul> <li><a href="https://ollama.com/library/gemma4" rel="nofollow">Gemma 4 QAT weights</a>: the Gemma 4 family is now optimized with Quantization-Aware Training (QAT) to dramatically reduce memory requirements and maximize on-device performance. Look for the tags ending in <code>-qat</code>: <ul> <li><code>gemma4:e2b-it-qat</code></li> <li><code>gemma4:e4b-it-qat</code></li> <li><code>gemma4:12b-it-qat</code></li> <li><code>gemma4:26b-a4b-it-qat</code></li> <li><code>gemma4:31b-it-qat</code></li> </ul> </li> </ul> <h2>What's Changed</h2> <ul> <li><code>ollama launch omp</code> now integrates with <a href="https://omp.sh" rel="nofollow">Oh My Pi</a>, an AI coding agent with IDE integration</li> <li>MLX embedding layers now use NVFP4 global scale for improved quantization on Apple Silicon</li> </ul> <p><strong>Full Changelog</strong>: <a class="commit-link" href="https://github.com/ollama/ollama/compare/v0.30.5...v0.30.6"><tt>v0.30.5...v0.30.6</tt></a></p>
<h2>What's Changed</h2> <ul> <li>Fixed the <code>gemma4:12b</code> floating point exception crash on x86, CUDA, Linux, and Windows systems.</li> <li><code>ollama launch hermes-desktop</code> now launches Hermes Desktop and can skip rebuilding when a packaged desktop app is already installed.</li> <li><code>ollama launch hermes</code> now supports native Windows installs through the Hermes PowerShell installer.</li> <li>Added Cline CLI integration docs.</li> </ul> <p><strong>Full Changelog</strong>: <a class="commit-link" href="https://github.com/ollama/ollama/compare/v0.30.4...v0.30.5"><tt>v0.30.4...v0.30.5</tt></a></p>
Developers & Open Source · vercel/next.js Releases
At Snowflake Summit 26 this week in San Francisco, the conversation moved into a new direction. If last year’s focus The post “Whoever builds the most joyous product wins”: The agent war begins appeared first on The New Stack . ]]>
“I’ve been doing this since the ’90s — decades of building guardrails to prevent humans from breaking production. We’re now The post Netlify CTO Dana Lawson: Writing code is no longer the job appeared first on The New Stack . ]]>
Sakana AI has launched a dedicated research lab for recursive self-improvement: AI that iteratively improves itself. The Japanese startup, co-founded by Transformer co-author Llion Jones, sees RSI as an alternative to the raw compute arms race among big US labs. Anthropic, meanwhile, warns about the control risks of this very technology. The article Sakana AI bets AI that improves itself can break the compute arms race of frontier labs appeared first on The Decoder . ]]>
Meta is developing a paid AI agent product called "Hatch" that could cost up to $200 per month. Users describe what they need in simple language, and Hatch builds working tools, schedules appointments, or sends emails. CEO Mark Zuckerberg sees the product as a way to open up new revenue streams beyond advertising and refinance the company's massive AI investments. The article Meta's Hatch AI agent could cost up to $200 a month and marks its first paid AI product appeared first on The Decoder . ]]>