标签：#ai

The Problem With Talking Directly to LLMs Most teams start by wiring their app straight to the OpenAI API. It works — until you need to add auth, rate limiting, observability, or swap out the model provider. Now you're rewriting application code instead of config. An AI Gateway solves this. One entry point, one place to govern traffic, providers become swappable. Kong Gateway is a mature choice here — it's been doing this for APIs for years, and the AI Proxy plugin extends that to LLMs. This post walks through the key ideas. For the full step-by-step guide, head over to the tutorial on Hashnode . What We're Building A Kong Gateway 3.14 data plane running on Kubernetes (kind locally), connected to a Kong Konnect control plane. The AI Proxy plugin sits on a route and handles forwarding to OpenAI — your app just talks to Kong. Your app → POST /ai/chat (Kong proxy) → AI Proxy plugin attaches API key → OpenAI API → response back to your app Your app never holds an OpenAI key. Kong does. You get rate limiting, logging, and model-swapping for free at the gateway layer. The Key Bit: decK Config as Code The most interesting part of this setup is using decK to define the service, route, and plugin as a YAML state file — then syncing it to Konnect, which pushes it down to the data plane automatically. # kong-ai.yaml _format_version : " 3.0" services : - name : openai-service url : https://api.openai.com routes : - name : openai-chat-route paths : - /ai/chat plugins : - name : ai-proxy config : route_type : llm/v1/chat auth : header_name : Authorization header_value : " Bearer $OPENAI_API_KEY" model : provider : openai name : gpt-4o options : max_tokens : 512 One sync command and Konnect pushes the config to every connected data plane: deck gateway sync kong-ai.yaml \ --konnect-token " $KONNECT_TOKEN " \ --konnect-control-plane-name "kong-ai-tutorial" Once it's live, a single HTTPie call confirms the whole chain is working: http POST localhost:8080/ai/chat \ Content-Type:applic

Reentrancy is the oldest trick in smart contract exploitation. The DAO fell to it in 2016. You would think a bug this famous would be extinct by now. It isn't. Protocols still lose millions to reentrancy every year, because the textbook version is the easy one, and attackers stopped using the textbook version a long time ago. This post walks through the classic bug, then the modern variants that still bite in 2026: cross-function, cross-contract, read-only reentrancy through view functions, and callback reentrancy through ERC777 and ERC721. I'll show why ReentrancyGuard is not the silver bullet most developers think it is, and how an LLM-based auditor reasons about the call-then-state-change pattern in a way pattern matchers can't. The Classic: Single-Function Reentrancy Here is the canonical vulnerable withdraw. contract VulnerableVault { mapping(address => uint256) public balances; function deposit() external payable { balances[msg.sender] += msg.value; } function withdraw() external { uint256 amount = balances[msg.sender]; (bool success, ) = msg.sender.call{value: amount}(""); require(success, "transfer failed"); balances[msg.sender] = 0; // state update AFTER the external call } } The attack is simple. The attacker is a contract. When withdraw sends ETH via msg.sender.call , control passes to the attacker's receive() function before balances[msg.sender] is zeroed. The attacker calls withdraw again. The balance is still the original amount, so the vault pays out again. Repeat until the vault is empty. contract Attacker { VulnerableVault vault; receive() external payable { if (address(vault).balance >= 1 ether) { vault.withdraw(); // re-enter before balance is zeroed } } } The fix is checks-effects-interactions. Do all your checks, then update all your state, and only then make the external call. function withdraw() external { uint256 amount = balances[msg.sender]; balances[msg.sender] = 0; // effect first (bool success, ) = msg.sender.call{value: amount}(""); //

Part 7 of 7 · Series: Building Your AI Developer Handbook · GitHub The Scenario You're building a password reset feature. User enters email → gets a reset link → clicks link → enters new password. Standard flow. Medium complexity. Let's walk through every step using the full workflow — as if you're looking over the shoulder of someone who built this system. "Show me your workflow and I'll show you your output quality." Before You Even Type Claude loads automatically in the background: ✓ ~/.claude/CLAUDE.md loaded ← the global handbook ✓ .claude/CLAUDE.md loaded ← project rules (TypeScript, pnpm) ✓ memory/MEMORY.md scanned ← all lessons and preferences You haven't typed anything yet. Claude already knows: Feature-based folder structure State management ladder No mocking the database No AI attribution in commits No useCallback without profiler evidence "A doctor who reviews your file before you enter the room is more useful than one who asks 'so, remind me who you are?'" Step 1: /status — Confirm the Setup /status Model: claude-sonnet-4-6 Effort: normal Plugins: security-guidance ✓ Thirty seconds. Sometimes the wrong model loads due to overload fallback. Sometimes a plugin fails silently. This check costs 30 seconds and prevents a surprise 30 minutes later. "A pilot's first action after sitting in the cockpit isn't to take off. It's to check all instruments are reading correctly." Step 2: /cost — Baseline /cost → Tokens used: 2,847 | Estimated cost: $ 0.004 Note this number. You'll compare it later before the expensive code review step. A surprise spike means something went wrong. Step 3: /plan — Design Before Coding /plan Build a password reset feature: - User enters email on /forgot-password - System sends a reset link (token, expires in 1 hour) - User clicks link → /reset-password?token=xxx - User enters new password - Token validated, password updated, token invalidated Claude responds with a plan — no code yet : Proposed approach: 1. DB: Add password_reset_tokens

Gemini in Chrome continues to roll out and has now landed in Latin America, the Middle East and Africa.

AI coding agent startup Niteshift has raised a $7 million seed round from a who's who of angels. It's betting companies will want power over, not lock-in with model makers.

Most of the value in SpaceX's IPO is effectively a call option on the company's ambitious space data center plans.

Through the acquisition, WMG aims to better track when its artists' work is used in AI-generated content or for training AI models.

Over 15 years after David Fincher and Aaron Sorkin's The Social Network premiered in theaters, Sony has released the first trailer for The Social Reckoning ahead of its theatrical release on October 9th. The follow-up takes place 17 years after the events portrayed in the original film and follows a young Facebook engineer who enlists […]

The funding round was led by Norwest, with participation S Capital VC, Cerca Partners, and Oceans Ventures. Snowflake Ventures also participated as a strategic investor.

Decart is launching Oasis 3, a real-time world model that generates photorealistic driving environments for autonomous vehicle testing, now available via API for developers to build on.

Insta360 has launched its first gimbal camera, the $770 Luna Ultra.

The generation side is shipping fast (TileGym, AutoKernel, KernelEvolve). The validation-side surface for “what the kernel actually did at runtime” has not kept pace. TL;DR In the past nine months, three significant releases have landed for auto-generation of CUDA kernels: NVIDIA TileGym , RightNow AutoKernel, and Meta’s KernelEvolve. Each ships training infrastructure for kernel generation. Validation infrastructure (what the generated kernel actually did at runtime, on a real workload, in a production-shaped environment) has not kept the same pace. eBPF traces are the ground-truth layer that closes the gap. What “validation” means at the kernel level Two distinct validation surfaces: Pre-launch: the generated CUDA C compiles, the PTX assembles, the kernel passes a numerical-equivalence test against a reference. Standard compiler / unit-test territory. Generation frameworks ship this themselves. Post-launch: the kernel ran, returned, took N microseconds, used M registers per thread, hit X cache miss rate, and did or did not serialize the rest of the stream behind it. This is the layer that an eBPF trace plus standard CUDA driver counters can answer for any kernel, generated or hand-written. Auto-generation pipelines do not by default close the post-launch loop. They demonstrate “the kernel works in our test setup”. They do not demonstrate “the kernel does not regress p99 latency on production inference traffic”. What an eBPF trace adds to a generated kernel Once a generated kernel is in a real workload, the same trace surface used for any CUDA kernel applies: launch latency from cudaLaunchKernel , sync stalls from cudaStreamSynchronize , host-side overhead from the dispatcher, host scheduling preemption while the GPU is busy. None of those signals are visible to a generation framework that evaluates kernels in isolation. -- post-launch validation: did the new generated kernel regress p99? SELECT kernel_name , COUNT ( * ) AS launches , AVG ( duration_ns ) / 1 e3 AS

Anthropic's strongest model is free until June 22 — and two more shifts for builders Three things landed for builders at once: the best model got cheaper (free, actually), free inference showed up on Apple's stack, and one still photo now becomes a talking video. Two of them you can act on right now. Here's the 90-second video version if you want the quick pass first: 1. Claude Fable 5 is public — and free on your plan until June 22 Anthropic released Claude Fable 5 , the first publicly available version of its Mythos-class model. It's state-of-the-art on nearly every benchmark Anthropic tests — software engineering, knowledge work, vision, and scientific research. It's free on Pro, Max, Team, and Enterprise plans through June 22 ; after that it's 10 dollars per million input tokens and 50 per million output . In high-risk areas (cyber, bio, chem) it refuses and falls back to Claude Opus 4.8 — about 95% of Fable sessions run entirely on Fable. This dropped just days after Anthropic publicly warned that AI was getting too dangerous. Why it matters: the strongest Claude is free to try on your existing plan for a two-week window. Run your hardest real task on it now and benchmark it before June 22 — the kind of jump that's worth re-checking your evals against. 2. Apple made its Foundation Models free for small developers At WWDC 2026 , Apple gave developers in the App Store Small Business Program (apps under 2 million first-time downloads ) free access to the next generation of Apple Foundation Models running on Private Cloud Compute — removing inference cost as a barrier. The Foundation Models framework now supports image input . A single Swift API can also call third-party models like Claude and Gemini, server-side. A new Dynamic Profiles system supports multi-agent workflows, and Apple will open-source the framework later this summer. Why it matters: you can ship AI features into an app without an inference bill. Prototype on Apple's free on-device models, and route

Uber's CTO told the world this month that the company spent its entire 2026 AI allocation by April. The story has been reported in a handful of outlets, hit the front page of Hacker News for 397 points and 469 comments , and is mostly being read as a cost-of-AI-tools story. It is one. It is also, on a closer reading of the numbers, a pricing-model story — and the structural fact that almost none of the coverage has emphasized is the one that determines whether this is a one-company anomaly or the beginning of an industry-wide budgetary crisis. The structural fact is that Claude Code, like most enterprise AI tooling in 2026, is priced on token consumption, not per-seat licensing. Token-based pricing scales with how aggressively the tool is used. Per-seat enterprise SaaS pricing — the model corporate IT budgets are built around — scales with how many people have access to it. Those two cost curves diverge in exactly the territory where productivity tools are designed to operate: high-engagement, daily-use, gradually-deepening workflows. The Uber data is the first public-facing version of a math problem most enterprise IT departments are about to discover privately. The numbers Uber CTO Praveen Neppalli Naga , named in Yahoo Finance's and Benzinga's coverage, said publicly that Uber is "back to the drawing board" on AI budgeting after the surge in Claude Code use blew through internal projections. The specific numbers, as reported across the multiple outlets covering the story: Claude Code adoption inside Uber's ~5,000-engineer organization went from 32% to 84% over four months. 70% of committed code at Uber is now AI-originated. 11% of live backend updates are "being written by AI agents built primarily with Claude Code," per the reporting. Per-engineer monthly API costs: $500 to $2,000. Uber's annual R&D spend is around $3.4 billion , of which the AI tooling line was a much larger fraction than expected. Cursor adoption plateaued; Claude Code dominated. These are ext

We train a model on handwritten digit classification. 99% accuracy . Then we train the same model on a new task — say, fashion item recognition. We go back and test it on digits. 34% accuracy . It has completely forgotten. Not gradually, not partially — almost entirely. What Just Happened? We trained a CNN on MNIST digits — 99.2% accuracy . After fine‑tuning on Fashion MNIST, it reached 91.1% accuracy . But when re‑evaluated on MNIST, accuracy collapsed to 33.9% . This collapse is catastrophic forgetting : the model’s weights shifted to optimize for the new task, erasing the old solution. Why did training on more data make the model worse at something it already knew? MNIST is handwritten digits (0–9). Fashion MNIST is clothing items like shirts and shoes. Both are 28×28 grayscale images, but the tasks are distinct. Why Does It Happen? The core issue is that the model relies on the same set of weights for both tasks. There is no separation or dedicated memory; every parameter is shared . When training shifts from Task A ( MNIST digits ) to Task B ( Fashion MNIST ), gradient descent simply minimizes the loss on the data it sees at that moment. It has no awareness that Task A ever existed. In the loss landscape, imagine two parabolic bowls: one for Task A and one for Task B. The optimum for Task A lies at θ A ∗ ​ , while Task B's optimum is at θ B ∗ ​ . As training on Task B progresses, the weights θ move towards θ B ∗ ​ . This movement inevitably raises the loss for Task A because its minimum is left behind. The root cause is the shared weight space. Gradient descent is a stateless optimizer; it only follows the current gradient signal. Since the minima for Task A and Task B are far apart, there is no single configuration of θ that satisfies both tasks simultaneously. This is why catastrophic forgetting occurs. Weight space can be visualized as an N-dimensional space, where each axis corresponds to one parameter. Every point in this space represents a full set of wei

OAuth for Remote MCP Servers How each AI assistant signs in to a remote MCP (Model Context Protocol) server, and why the flow differs by client and by where it runs. Overview The protocol throughout is standard OAuth 2.1 — an open, widely implemented authorization standard. The human sign-in runs through oauth2-proxy , one of the most widely deployed open-source auth proxies; the only deployment-specific piece is a thin, spec-conforming authorization server (the /oauth endpoints) that hands MCP clients their tokens. Every client ends up the same way — a person signs in against Google (restricted to your organization's domain), and the client holds a short-lived bearer token it presents on each /mcp call. Two things differ between assistants: where the client runs (a machine on the VPN — private — vs. the vendor's cloud — public ), which decides the host it reaches; and what kind of OAuth client it is — a public client proving itself with PKCE (Proof Key for Code Exchange, which lets a client with no secret prove the token request comes from the same client that started the flow), or a confidential client proving itself with a secret. The participants oauth2-proxy — the public-facing reverse proxy. It authenticates the human against Google (the sign-in restricted to your organization's domain) and forwards the verified identity to the app behind it. Only oauth2-proxy faces the internet. It is a mature, heavily-deployed open-source project — the standard way to put Google/OIDC (OpenID Connect) single sign-on in front of a service, widely used in Kubernetes deployments — so the most security-sensitive leg of the flow (the OAuth exchange with the identity provider) runs on battle-tested code. The MCP server — the app on a loopback port behind the proxy. It plays two roles: the OAuth authorization server ( /oauth/authorize , /oauth/token , /oauth/register , .well-known discovery) and the /mcp tool endpoint. It mints codes and tokens, and validates a token on every /mcp c

I spent a weekend wiring Google's Gemini and Veo APIs into a single app just to feel where the edges of multimodal AI actually are. It turned into a small studio I now use daily, and along the way I learned more about these models from plumbing them than from any paper. Here's the honest technical debrief. Three pipelines, three completely different problems I wanted one prompt box that could do video, image editing, and document Q&A. Naively I assumed they'd share most of the stack. They don't. 1. Image-to-video: the enemy is time, not pixels Generating one good frame is solved. Video is about temporal coherence — frame 13 must agree with frame 12 or you get flicker and identity drift. Modern video models treat the clip as one object in space and time (latent diffusion over a width x height x time volume, with spatiotemporal attention) rather than 120 independent images. Conditioning on a reference image as the first frame is what makes image-to-video feel controlled: you've handed the model a strong anchor and asked it to extrapolate motion, not invent a world. The surprise: native audio sync (Veo 3.1 generating clip + soundtrack jointly) does more for perceived realism than another notch of resolution. A door slam landing on the exact frame the door shuts is uncanny in a good way. 2. Instruction-based image editing: preservation is the hard part Generating is unconstrained; editing must change one thing and preserve everything else. Condition the diffusion model on both the instruction and the source image's latents, cross-attend the instruction to steer only the referenced region, and bias hard toward preserving unedited latents. Push that preservation too soft and the subject's face quietly morphs across edits — the classic 'character consistency' failure that makes or breaks storytelling use-cases. 3. PDF chat: it's retrieval, not a long context The naive 'paste the whole PDF' approach dies on long files (models get lost in the middle ) and costs you the full

What device do we use almost all the time? Our mobile phone, almost certainly. If we wanted to...

Adi Polak discusses the architecture required to transition from stateless prompts to state-aware, context-rich AI agents. Drawing on 15 years in distributed systems, she shares how engineering leaders can leverage Apache Kafka and Flink for real-time stream processing, dynamic memory tiering, and tool orchestration via MCP to solve token limits, cost spikes, and latency bottlenecks. By Adi Polak

A new report from OpenAI details PRC-linked influence operations using AI to target U.S. tech debates, data center narratives, tariffs, and false claims about ChatGPT.