AI News Today

I run GrowthSpree, a B2B SaaS marketing agency, selling into the US and Europe. Six months ago we made a bet that the agency model itself was broken for the AI era, so we tore ours down and rebuilt it. This is what we did and what actually happened. The bet: become the first truly AI-native marketing agency. Not an agency that "uses AI tools" by bolting ChatGPT onto a 2015 workflow, but one where AI is wired into how we find demand, win clients, and run delivery. We rebuilt the operating model,

<!-- SC_OFF --><div class="md"><p>​</p> <p>For open-weight LLMs, how practical is it to study defenses against post-release fine-tuning that weakens refusal or safety behavior?</p> <p>I've been seeing “uncensored” or “heretic” variants of new models appear very quickly after release, which raises a question I’m curious about: is fine-tuning resistance a meaningful safety goal for open-weight releases, or is it too narrow because determined users can always modify weights, switch models, or use o

As the parent of two little girls, I often think about how their childhood is different from mine. The seven-year-old is learning about AI at school. The five-year-old is given internet-based homework every week. And they are both absolutely repulsed by the idea of smoking. That was not the prevailing sentiment when I was young.…

Grill, mow, and stay cool this weekend while saving hundreds at Lowe's 4th of July sale.

Kuaishou has raised about $2 billion from investors for its AI video division, Kling. The article Chinese AI video maker Kling raises $2 billion as it gears up for Hong Kong IPO appeared first on The Decoder .

AI agents will happily create 1000+ line source files and add a 20th parameter to a function call, even if there's a rule file telling them not to. So I built Scopewalker: a local MCP server (open source, runs over stdio, and makes no network calls) that gives agents actual numbers to check against. It gives the agent 8 read-only tools: line counts, cognitive complexity/nesting/parameter counts, an oversized-file/function checker against configurable thresholds, doc coverage, code smell detectio

Comments

Article URL: https://github.com/gfernandf/agent-skills/blob/main/docs/ACTION_PREFLIGHT_FORECAST_QUICKSTART.md Comments URL: https://news.ycombinator.com/item?id=48772302 Points: 1 # Comments: 2

A sea of quantum weirdness. ScienceAlert stories are written, fact-checked, and edited by humans, never generated by AI. Don't miss a story, subscribe here.

Article URL: https://digitvest.com/en Comments URL: https://news.ycombinator.com/item?id=48772198 Points: 1 # Comments: 0

Article URL: https://medium.com/@gianlucabailo/the-day-i-played-hide-and-seek-with-an-ai-7a1ee189b5ff Comments URL: https://news.ycombinator.com/item?id=48771783 Points: 2 # Comments: 1

In "Silo," the characters speak to a mysterious voice known as "the algorithm." But have we already met the voice's owner?

Article URL: https://www.youtube.com/watch?v=I9F_VFfLTmM Comments URL: https://news.ycombinator.com/item?id=48771721 Points: 2 # Comments: 1

<h2> Introduction: The Need for Modern Web/AppSec Training </h2> <p>The cybersecurity landscape is evolving at a breakneck pace, but the tools we use to train the next generation of defenders are stuck in the past. Most web/appsec learning platforms still focus on <strong>basic, textbook vulnerabilities</strong>—XSS popups, simple SQL injection, or trivial IDORs. These labs are like teaching someone to swim in a kiddie pool; they might grasp the concept, but they’re <em>ill-prepared for the open

Article URL: https://github.com/anliberant/obsidian-ai-setup Comments URL: https://news.ycombinator.com/item?id=48771700 Points: 1 # Comments: 0

Hi HN,I am building opplic.com an AI employee that you can hire for each of your client. The idea is simple, if you are an agency and managing 10 clients it becomes hectic to manage all if you are doing alone, you might need to hire a SEO guy but wait, 90k per anuum for a single guy !? Think if you want to hire one for each that's almost a million. Here's where opplic comes into picture, Instead of generating random marketing ideas, Opplic continuously learns each client's website, business, ser

<p>Over the past year, AI has become part of many developers' daily workflow. It can generate code, explain unfamiliar frameworks, review pull requests, and even suggest architectural patterns.</p> <p>But I've noticed that the biggest impact isn't on writing code faster. It's on how we think about software architecture.</p> <p>With AI handling repetitive implementation tasks, it feels like architects and senior engineers are spending more time on system design, scalability, security, integration

I’ve been working on this for just a year, mostly with hand coding. I only started using agents in the last few months. It’s not fully finished yet, but the waitlist is open now. I’m giving 80% off for the first three months to everyone who joins the waitlist. Really hope to see your reviews and hear what you think! Comments URL: https://news.ycombinator.com/item?id=48771562 Points: 1 # Comments: 0

<h2> The trust problem nobody scopes correctly </h2> <p>When companies talk about trust in AI, they almost always mean trust in the model. Is the output accurate? Is it hallucinating? Can we rely on what it says?</p> <p>Those are valid questions but they're the wrong starting point. The trust that actually determines whether AI gets adopted or quietly abandoned inside an organization isn't about the model. It's about the system surrounding it.</p> <h2> The four questions that determine </h2> <p>

I'm a bit annoyed by the feeling that we're kind of stuck when it comes to using LLMs for programming. I use Claude Code and Codex, but I haven't been able to enter flow state like I can when I hand write code. This is kind of ironic to me since AI should be a bicycle for the mind, but right now it feels like a bicycle that just brakes abruptly every couple minutes. I stop, wait, review, prompt again. Is there anyone exploring something fundamentally different than the prompt response loop we ha

Article URL: https://www.bbc.com/news/articles/cgrkd41n2v9o Comments URL: https://news.ycombinator.com/item?id=48771466 Points: 3 # Comments: 1

Comments URL: https://news.ycombinator.com/item?id=48771421 Points: 1 # Comments: 1

Article URL: https://soumitradutta.gt.tc Comments URL: https://news.ycombinator.com/item?id=48771271 Points: 1 # Comments: 0

Article URL: https://www.youtube.com/watch?v=59XdUJRzRxc Comments URL: https://news.ycombinator.com/item?id=48771229 Points: 1 # Comments: 0

Article URL: https://agentrc.ai/ Comments URL: https://news.ycombinator.com/item?id=48771202 Points: 1 # Comments: 0

Article URL: https://alprado.com/blog/i-still-enjoy-building-websites-without-ai/ Comments URL: https://news.ycombinator.com/item?id=48771003 Points: 4 # Comments: 1

I'm a software engineer and have been working on this for a couple of months, for fun, and thought I'd share it. If you try it, let me know, I'd be stoked! I was inspired by a friend who wanted to escape the algorithms and had built a small, private reader for himself. He wasn't sure he'd ever release it, so I built my own. It's yet another self-hosted RSS reader, I know. I've been trying to build things like inline YouTube (with filters to drop Shorts and live streams) and webhooks into the cor

Article URL: https://arxiv.org/abs/2606.22737 Comments URL: https://news.ycombinator.com/item?id=48770893 Points: 4 # Comments: 0

I am trying to setup Mr.Jassy as a Slackbot app. I want Mr. Jassy to be our personal AWS butler. Mr.Jassy has access to AWS resources in read only and can help the team answer questions. The way its setup now is that I run 'claude -p' from inside a lambda and is invoked by slack events. Is this the right approach or is there a better way to do this? I also maintain thread context by resuming a previous session. Comments URL: https://news.ycombinator.com/item?id=48770691 Points: 2 # Comments: 0

Article URL: https://www.fastcompany.com/91566861/how-to-avoid-ai-in-as-many-places-as-possible Comments URL: https://news.ycombinator.com/item?id=48770606 Points: 6 # Comments: 1

arXiv:2607.01235v1 Announce Type: new Abstract: Understanding how Large Language Models (LLMs) make token-level decisions during code generation remains a major challenge for both researchers and practitioners. While recent tools provide insights into model internals or generation outcomes, they often lack decoding-time signals, fine-grained uncertainty measures, and interactive mechanisms for exploring alternative generation paths. We present TokenScope, an interactive interpretability and anal

arXiv:2607.01236v1 Announce Type: new Abstract: As LLM agents gain increasing access to powerful tools, ensuring that their actions are aligned with the user's intent becomes critical. When an agent's proposed tool invocation deviates from the user's intent -- a phenomenon called misalignment -- it may lead to harmful consequences that are difficult to undo. Existing runtime guardrails rely on an LLM-as-a-judge paradigm that lacks a systematic framework for reasoning about alignment, often produ

arXiv:2607.01237v1 Announce Type: new Abstract: Reasoning language models often generate long chain-of-thought (CoT), which accumulates a massive KV cache during the decoding phase and incurs high decoding latency and limited throughput. To address these issues, KV cache compression has emerged as a promising technique for reducing memory overhead by selectively removing unimportant KV pairs while preserving useful ones for subsequent decoding. Nevertheless, we identify two key limitations in ex

arXiv:2607.01239v1 Announce Type: new Abstract: Character-level perturbations bypass safety alignment in modern LLMs despite leaving prompts human-readable. We identify and test a central structural mechanism: BPE tokenization fragments safety-critical words into sub-word pieces, and the three public alignment datasets we surveyed contain no intentionally fragmented inputs. The mechanism is a chain, tested end-to-end on five model families (Qwen-3-4B, Qwen-2.5-7B, Gemma-3-4B, Llama-3.1-8B, Mistr

arXiv:2607.01240v1 Announce Type: new Abstract: Count-based F1 is widely used as a proxy for LLM error-detection quality, but this paper shows that it can rise dramatically without a corresponding improvement in span localization, a gap termed F1 Inflation. The paper introduces ErrorBench, a controlled stress-test protocol for prompt-induced count distortion. ErrorBench evaluates six contemporary LLMs under five prompt conditions over 4,290 responses from 143 CoNLL-2014 passages. Under CoNLL-201

arXiv:2607.01241v1 Announce Type: new Abstract: Existing prompt compression methods treat text as flat token sequences, failing to capture the distributed nature of important information, which is often spread across multiple locations and connected through both local syntactic dependencies and global semantic relations. Such relational structure is naturally represented as a graph, where tokens or sentences become nodes and their dependencies become edges. To this end, we propose RAGP, which fo

arXiv:2607.01245v1 Announce Type: new Abstract: We introduce Office Comprehension Bench (OCB), the first public benchmark to jointly evaluate LLM systems on Word, Excel, and PowerPoint comprehension over native file formats (.docx, .xlsx, .pptx) and their variants. OCB consists of two tracks. File Fidelity Q&A tests structural and visual perception of office artifacts - tables, charts, embedded images, formulas, and app-specific elements such as headers, speaker notes, and named ranges. Domain Q

arXiv:2607.01293v1 Announce Type: new Abstract: We present RuleChef, a framework that uses large language models (LLMs) to generate executable rules for NLP tasks such as text classification, Named Entity Recognition (NER), or relation extraction. Rules are generated based on a task description and a set of labeled examples, then they are iteratively improved based both on additional examples and on human feedback overexisting rules. RuleChef can also be used to bootstrap rules using the observe

arXiv:2607.01392v1 Announce Type: new Abstract: Aligning large language models with diverse and heterogeneous human values requires multi-objective alignment methods to effectively trade off conflicting preference dimensions. Current methods achieve this trade-off by training policies conditioned on preference vectors and leveraging online direct preference optimization. However, exploration uncertainty can cause the reward distributions of responses generated under different preference vectors

arXiv:2607.01420v1 Announce Type: new Abstract: As grounded QA systems are increasingly deployed in AI assistants, accurately attributing generated answers to evidence is critical for user trust and model safety. While unimodal attributions have been explored in depth, the multimodal setting remains relatively under-researched. As a result, we introduce MultAttnAttrib, a training-free attribution-generation method that leverages a model's prefill pass, selected attention heads, and calibrated th