|
|
Hi there, this is your daily ☕️ DevShot.
|
|
|
In today's DevShot:
|
|
🤖 Google delays Gemini 3.5 Pro over coding 🇨🇳 China's Kimi K3 rivals top AI models 🔑 Claude can now use your 1Password logins 🐧 AMD driver slows Linux GPUs 42x 📺 Netflix drops gRPC for SSE Plus: 🎁 6 other news you might like, 🧰 6 tools, and 📚 5 papers.
|
|
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
divider under the main Beehiiv ad removed per request
|
🤖 Google delays Gemini 3.5 Pro over coding
LINK
|
- Google has pushed back the launch of Gemini 3.5 Pro past its promised June window, citing the need to sharpen the model's coding performance before releasing it to partners and developers.
- A late-June attempt to boost coding skills by updating Gemini's training data reportedly produced disappointing results, suggesting a development reset between the I/O 2026 reveal and the missed deadline; Gemini 3.1 Pro still dates to February.
- Google says it is currently testing 3.5 Pro, an upgraded Flash model, and others with partners, while internal engineers hit AI capacity limits and an effort to consolidate the company's separate coding tools moves forward.
|
🇨🇳 China's Kimi K3 rivals top AI models
LINK
|
- Moonshot AI released Kimi K3, a 2.8-trillion-parameter model available now via its website and API, with an open weight release promised by July 27, 2026, surpassing DeepSeek's 1.6T v4 Pro in size.
- Priced at $3 per million input tokens and $15 per million output, K3 matches Claude Sonnet and doubles the cost of the prior Kimi K2.6, while using ~20% fewer output tokens on the Artificial Analysis Intelligence Index.
- Self-reported benchmarks show K3 beating Claude Opus 4.8 and GPT-5.5, trailing Claude Fable 5 and GPT-5.6 Sol, and topping Arena.ai's Frontend Code leaderboard; it currently ships with only a single "max" reasoning effort.
|
🔑 Claude can now use your 1Password logins
LINK
|
- 1Password has launched a browser integration for Claude that lets Anthropic's agent use your stored logins to complete multi-step tasks like booking travel and managing accounts without you manually entering credentials.
- A "zero-exposure security framework" injects the needed credentials through a secure channel Claude can't read, so the agent never sees your passwords or MFA one-time codes, and 1Password scans each page after autofill.
- Access is granted per task via a single biometric prompt, with the vault locking down to only explicitly approved credentials; it's live now on Mac across all plans, with payment cards and identity details coming later.
|
🐧 AMD driver slows Linux GPUs 42x
LINK
|
- An AMD GPU driver bug on Linux is causing severe performance regressions of up to 42x, with the Ubuntu kernel team warning users across all major distributions that the slowdown affects many graphics workloads right now.
- The regression hits AMD GPU workloads broadly on current kernels and was uncovered through Phoronix Test Suite benchmarking, and the Ubuntu kernel team has flagged it as a temporary issue that is being actively addressed upstream.
- Linux users running AMD GPUs should track Phoronix and the Ubuntu kernel mailing list for the upstream fix, and consider pinning to a known-good kernel until a patched driver ships to their distribution.
|
📺 Netflix drops gRPC for SSE
LINK
|
- Netflix has replaced the gRPC framework with the Server-Sent Events protocol to expose the real-time service dependency map its engineers use to troubleshoot issues across the company's increasingly complex streaming, gaming, and ads stack.
- The team found gRPC became a performance bottleneck, where serialization overhead, connection pool management, and memory pressure for streaming responses consumed more CPU than business logic while message queues added infrastructure complexity without benefit.
- Switching to SSE gave lighter serialization, natural backpressure with reactive streams, and a simpler connection model, evening out CPU usage across instances to eliminate hot nodes carrying 10x the load and cutting network bandwidth significantly.
|
|
📚 Tutorials & Advice
|
> Octopus Easy Mode - Argo CD Manifest Update: Set up a practice Octopus project with mocked Git and Argo CD so teams can learn how deployments commit environment-specific config changes without needing a real Kubernetes cluster.
|
|
|
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
🔗 Other News & Articles
-
How Uber Builds Zone-Failure-Resilient OpenSearch Clusters
LINK
-
Firefox in WebAssembly
LINK
-
LM Studio expands beyond chat with Bionic, a new AI agent app for open models
LINK
-
GoDaddy opened its registrar to AI agents. Then it had to build guardrails.
LINK
-
Roll it out, roll it back, never redeploy
LINK
-
Microsoft open-sources Comic Chat, the software that introduced Comic Sans to the world
LINK
|
|
📄 Papers & Resources
|
> Humanoid robot control gets a scaling recipe that combines more training rollouts, more varied reference motions, and a new model design, producing robots that move more naturally and generalize better in both simulation and real-world tests.
LINK
|
|
> Home safety scans get a real world test kit of 66,668 question and answer pairs, and training on it markedly improves how well vision AI models spot hazards like fire or fall risks in actual rooms, not just simulated ones.
LINK
|
|
> Robot decision-making under uncertainty now learns its own world model from simulated data instead of hand-built rules, matching or beating a leading benchmark method on under 10% of the training data while cutting planning costs nearly a thousandfold.
LINK
|
|
> Uncertainty scoring in AI now has a unified formula, deriving a model's confidence versus true unpredictability from the loss function it already uses, replacing today's patchwork of incompatible measures across the industry.
LINK
|
|
> Synthetic data for causal analysis can look realistic and predict outcomes well while still badly distorting treatment-effect estimates, so a hybrid method that models treatment and outcome separately preserves accuracy for real decision-making.
LINK
|
|
|
|
🧰 Latest Tools
|
Zro: runs open-source coding models like MiniMax M3 and GLM-5.2 via one endpoint on EU infrastructure, with zero request retention.
LINK
|
|
Codex Micro: a platform for building and scaling AI-powered products, offering access to leading models and developer tools.
LINK
|
|
Nitrosend: lets AI agents self-onboard to send marketing, transactional, and customer-reply emails from your own domain, with human approval built in.
LINK
|
|
Mojibake: a low-level Unicode library written in C, implementing normalization, case conversion, segmentation, bidirectional text, collation, and confusable detection.
LINK
|
|
Captchainbox: makes unrecognized email senders solve a captcha or pay a fee before their message reaches your inbox, filtering AI-generated spam automatically.
LINK
|
|
linux-stable-msm-dtsi: a modern Linux kernel port enabling a ten-year-old HTC QWERTY phone to function as a handheld terminal device.
LINK
|
|
|
|
💬 How did you find today's edition?
We read every reply — just reply to this email and let us know how we can improve!
|
|
