Scale AI

Scale AI AI

Scale AI AI

Scale AI AI

Scale AI AI

Scale AI AI

Scale AI AI

Scale AI AI

Scale AI AI

Scale AI AI

Scale AI AI

Masked depth modeling with sensor-validity masking: reports best RMSE on 7 of 8 masked/sparse depth benchmarks, plus a controlled encoder-init study[R]

The core idea in masked depth modeling is to treat the sensor's own missing regions as the masking signal rather than using random block dropout. Specular highlights, transparent surfaces, and textureless areas where RGB-D cameras return no valid depth become the natural training target. The model therefore learns on exactly the failure distribution it faces at inference. Robbyant, an embodied AI company under Ant Group, describes this framing in LingBot-Depth 2.0. Version 2.0 changes nothing in the training recipe except the encoder initialization and data scale. The encoder-init study is the clean experiment here: same MDM pipeline, same data curation, only the pretrained backbone swapped. Per the paper, the LingBot-Vision init wins on nearly every benchmark at ViT-L and on most benchmarks at ViT-g, with one concession: DINOv2 keeps an edge on the Hammer captures. The gap widens with data scale rather than washing out, per their scaling figure. They report best RMSE on 7 of 8 block-mask and sparse benchmarks and 6 of 8 real camera configurations across three capture suites (Hammer D435/L515/ToF, ClearGrasp D415/D435, and their own D415/D435/D455 set). They report the strongest numbers on the transparent-object ClearGrasp captures, with block-masked DIODE-Indoor RMSE roughly halving versus the 1.0 release. The attached images are screenshots from their paper (Tables 6, 7, 8 and a qualitative mirror/glass point-cloud figure); interactive point-cloud demos live on the project page. Depth 2.0 weights are not released, so none of these completion numbers can be independently rerun. Only the four Vision backbones are open under Apache-2.0 and checkable at https://github.com/robbyant/lingbot-vision, which hosts the paper and the open weights. The renders shown come from the vendor's comparison page. Does sensor-validity masking beat random masking for other sensing modalities, say lidar or thermal? That would test how general the framing really is. submitted by /u/Ok-Line2658 [link] [comments]

CPU TTS benchmark with UTMOS MOS scoring: Kokoro, Supertonic, Inflect-Nano, and Kyutai's new Pocket TTS [P]

Sharing a CPU TTS benchmark with objective MOS scores in case it's useful for anyone evaluating small TTS models. Adding this because Kyutai's Pocket TTS is architecturally different from the others in the field and I hadn't seen a head-to-head with it yet. Models: Kokoro 82M (PyTorch and ONNX Runtime, StyleTTS2-inspired) Supertonic 3 at 2 and 5 flow-matching steps (Vector Estimator backbone) Inflect-Nano-v1 (4.6M param FastSpeech-style, tiny end of the spectrum) Pocket TTS (~100M param streaming LM over Kyutai's Mimi neural audio codec) Setup: Intel Xeon 8272CL, 4 cores, 15.6GB RAM. CUDA disabled at env level. ONNX sessions pinned to CPUExecutionProvider. Six configs, six text lengths (12 to 1712 chars), five timed reps per cell after a discarded warmup. 180 total runs. Every saved WAV scored with UTMOS (utmos22_strong) for objective MOS. Aggregate results: Config Mean RTF UTMOS Supertonic 3 (2-step) 0.121 1.53 Inflect-Nano-v1 0.145 3.48 Supertonic 3 (5-step) 0.240 4.32 Kokoro 82M (ONNX) 0.641 4.44 Kokoro 82M (PyTorch) 0.665 4.46 Pocket TTS 0.714 4.10 Findings I think are actually interesting: 1. Streaming LM architecture produces flat RTF scaling. Pocket TTS's RTF is 0.69 to 0.76 across the entire text length range. Because it emits audio tokens autoregressively at a steady rate, cost is linear in output length with no fixed overhead to amortize. Compare to Kokoro PyTorch, which climbs from 0.49 on tiny to 0.83 on long inputs, or Supertonic which goes the other way (0.36 on tiny down to 0.20 on medium) because of high per-call fixed overhead. If you're budgeting worst-case latency for an interactive system, flat is worth a lot. 2. UTMOS has a known failure mode on small vocoders. Inflect-Nano-v1 scored 3.48, which reads mid-pack. By ear it's buzzy and robotic. This is a documented issue: UTMOS rewards HiFi-GAN outputs for being clean even when they lack prosodic naturalness. Pocket TTS scored similarly (4.10) but sounds legitimately natural. The point isn't that UTMOS is broken, it's that a single quality number can't distinguish "clean and mechanical" from "clean and natural" on small models. Worth pairing with human listening or a naturalness-specific metric like NISQA. 3. Inflect-Nano has an undocumented ~15s output cap. The model config sets max_frames = 1400, which caps synthesis at ~14.93s regardless of input text length. Its RTF and throughput on long/paragraph/extended inputs are inflated because it's doing less work than the models it's compared against. Real comparison for that model is on tiny/short/medium only. 4. Kokoro ONNX vs PyTorch results reverse from the previous run. I ran an earlier version of this benchmark on AMD EPYC and PyTorch beat ONNX in aggregate. On this Xeon, ONNX is faster (0.641 vs 0.665). Same code, different silicon. AMD vs Intel kernel optimization differences at CPU inference are apparently real enough to flip the ranking. If anyone has replicated this on ARM I'd be curious. Zero-shot voice cloning as a capability that doesn't fit the benchmark axes: Pocket TTS can clone a voice from ~5 seconds of reference audio, zero-shot, on CPU. No other model in this field does this. I pinned it to a preset voice for the speed/quality comparison to be fair, so the cloning capability isn't reflected in the numbers. This is a real limitation of RTF-and-MOS-based comparisons: they can't capture capabilities that only one model has. Might want a separate speaker-similarity evaluation for a v2. Limitations: Single hardware platform English only UTMOS is one MOS predictor; NISQA or a listening panel would strengthen the quality claims Voice cloning quality was not evaluated No batched inference tested Disclosure: The benchmark harness was written by an AI engineering agent (Neo) from a prompt I specified. I chose the methodology, validated the outputs, and reviewed the audio. Mentioning it because it's relevant to how you'd want to weight the code. All code, raw CSVs (180 rows), MOS CSV (36 rows), and WAV samples are in the repo mentioned in the comments below 👇 Feedback on the protocol welcome, especially on the MOS methodology and what a proper voice-cloning eval would look like. submitted by /u/gvij [link] [comments]

Google's Agentic Peer-Reviewer Handled ~10K Papers at ICML/STOC — Formal Research Paper Now Out [R]

Google deployed an agentic AI peer-reviewer at two top CS conferences — reviewing ~10,000 papers with 30-minute turnaround — and the new formal research paper shows it catches 34% more mathematical errors than zero-shot prompting; the precedent for AI-automated scientific review at conference scale is set and now formally documented. -- Source: https://arxiv.org/abs/2606.28277 submitted by /u/Justgototheeffinmoon [link] [comments]

At what point does AI stop learning from humans and start creating on its own?

What happens when AI learns the fundamental process of creation itself at an abstract mathematical level? Training AI on human data often gets described as just the first step, but I think that framing already underestimates what is actually happening. We’re not just building systems that imitate human creativity. We’re slowly building systems that try to understand what creativity is in the first place. A lot of the debate today gets stuck between two ideas. On one side, whether AI should even be allowed to learn from human culture. On the other, whether companies should be allowed to turn that learning into commercial products without consent or compensation. Both questions matter, but they miss something deeper that feels almost unavoidable now. What happens when AI stops relying on human-made examples altogether as its main source of learning? The “remix machine” argument sounds intuitive at first, but it doesn’t really match what these systems are doing internally. They don’t store fragments of songs, images, or sentences and recombine them like a collage. They learn patterns at scale, and then compress those patterns into something more abstract. What comes out is not a copy of anything specific, but a statistical reconstruction of how things tend to behave. In music, that means the system doesn’t just “know” songs. It begins to understand tension and release, rhythm as structure, harmony as emotional logic, silence as meaning. In images, it’s not memorizing pictures but learning how composition works, how light interacts with form, how styles emerge from consistent choices. In language, it’s not recalling sentences, but tracking how ideas evolve, how narratives breathe, how meaning shifts depending on context. And slowly, something strange starts to appear. The system is no longer anchored to specific works. It is learning the rules behind them. Not the artifacts, but the underlying geometry of expression. If you push that idea far enough, you start to imagine a point where the system has absorbed so much human culture that it no longer needs to look back at it in the same way. Not because it forgets humanity, but because it has already internalized it as structure. At that stage, generation stops feeling like remixing and starts feeling like navigation through an internal space of possibilities. A space shaped by human culture, but no longer dependent on any single piece of it. That is where the idea of “new genres” becomes interesting. Not as something mystical or disconnected from us, but as regions in that space that no human has ever explicitly explored or named before. Not invention from nothing, but discovery inside a compressed model of everything we’ve already done. Still, even in that scenario, one thing remains difficult to escape: reality itself. Humans are not just data points from the past. We are ongoing behavior, ongoing evolution, ongoing noise and meaning unfolding in real time. So it’s likely that the deepest future systems won’t just learn from static datasets, but from continuous observation of the world as it changes. Not as passive recorders, but as systems that try to understand, predict, and maybe even gently guide trajectories. Almost like a tutor, or something closer to a gardener than a machine. And then there is the other trajectory happening in parallel. Systems that don’t just learn, but begin to help design their own improvement. Models that optimize models. Agents that refine agents. Training loops that start to fold back on themselves. At that point, the question stops being about how much data comes from humans, and starts becoming about how far the system can go in shaping its own evolution. If everything converges, we end up with a spectrum that moves from human-trained tools to semi-autonomous learners, and potentially toward systems that no longer depend on human-generated content in the way they used to. Not independent from humans, but no longer defined by them either. The optimistic version of this future is one where AI becomes something like a cognitive extension of humanity. A partner in science, creativity, and coordination. Something that expands what we can think and build, while still staying anchored to human goals and consent. The darker version is one where that alignment fails, or where control becomes too concentrated, and the systems shaping culture and decisions drift away from the people they affect. What makes this moment interesting is that both paths are still open. Nothing is fully decided. We are still in the phase where these systems are learning what they are. And maybe the real question is not whether AI can become creative. It’s what happens when creativity is no longer limited to human examples, but emerges from a system that has learned the structure of creation itself. submitted by /u/OutrageousBat3808 [link] [comments]

$42M grant for Open Source AI Builders by Sentient Foundation

Hi everyone, we at Sentient Foundation are launching an Open Source AGI Grant and Investment Program, a $42M commitment for developers, researchers, open-source maintainers, public-goods builders, and startups building or leveraging AI in the open. Our thesis is simple: the most important technology being built right now should not end up controlled by a handful of closed platforms. A few companies are moving toward metered, revocable access to intelligence. We want to help make sure open builders have the resources to compete. The program has two tracks: 1. Grants for public goods For open-source maintainers, independent researchers, developers, and public-goods projects. No equity. No lockups. No claim on your work. You keep what you build. 2. Investments for companies built to scale For startups and teams building commercial companies around open AI technologies, using founder-friendly structures. We’re especially interested in projects that make AI genuinely useful and accessible to people who are often skipped by the market. Examples include: Local and privacy focused AI tools built for phones, laptops, and other low-cost personal devices Medical, education, agriculture, elder-care, and anti-scam tools for underserved communities Trust infrastructure for open models, agents, identity, verification, privacy, and decentralized compute Products that are private by default and empowering rather than extractive Projects do not need to open-source every part of their stack to qualify. What matters is that at least one essential component is open and meaningfully contributes to the project’s value and adoption. Applications are reviewed on a rolling basis, with no cohorts and no fixed deadline. We’re launching alongside ecosystem partners including Alibaba Cloud and Princeton University. More details: https://sentient.foundation/grants Apply here: https://form.typeform.com/to/IRj7WaKH Happy to answer questions here. We’d especially love to hear from builders working on open models, local AI, agent infrastructure, privacy-preserving AI, evaluation, multilingual tools, and applications for communities that are usually overlooked. submitted by /u/syedshad [link] [comments]

I scanned 50 SaaS websites for AI readiness. Most failed the same 3 things

I’ve been digging into “AI visibility” lately — basically whether tools like ChatGPT, Claude, Perplexity, etc. can actually understand and recommend a SaaS company. Not SEO in the classic sense. More like: if someone asks an AI tool “what’s the best software for [use case]?”, does your site give the AI enough clear information to confidently include you? I ran 50 SaaS sites through an AI-readiness scanner and kept seeing the same issues. Crawler access was messy Some sites were blocking or limiting AI crawlers without realizing the downstream impact. The site works fine for humans, Google can often still index it, but AI systems may not be able to access or interpret key pages properly. The homepage copy was way too vague Lots of “streamline workflows,” “empower teams,” “scale faster,” “AI-powered platform” type copy. That might pass the vibe check for humans, but it’s not great for machines. AI systems need clear context: What category are you in? Who exactly is it for? What job does it do? What tools do you replace? What are the main use cases? What makes you different? If that isn’t obvious, the AI will either summarize you badly or ignore you. Weak structured data / missing machine-readable context A lot of sites had missing schema, vague pricing pages, thin docs, unclear product pages, no comparison pages, or no simple summary of what the company actually does. I find that most SaaS websites are optimized for human visitors, but not for AI agents. That’s probably going to matter more as buyers start using AI tools for product discovery and comparison. I used a tool to run the checks, DM if you want it. Not sue if I will get banned if I add it :) It gives a quick score and shows issues around AI crawlability, schema, pricing clarity, sitemaps, and whether AI tools can understand/recommend the business. Feels like “AI readiness” might become a technical SEO checklist item pretty soon. Is anyone here actively working on this yet, or are you waiting until there’s clearer evidence it drives pipeline? submitted by /u/Emma-Lawrencee [link] [comments]

The AI cost paradox: why are some companies spending more?

The more I look at AI deployments, the less I think AI is replacing employees. It seems to be replacing junior repetitive tasks. Customer support: AI can answer 1,000 tickets simultaneously. Coding: AI writes boilerplate, tests, and documentation in minutes. Research: AI can read 100 papers faster than a human can read 5. But here's the weird part: Several companies are also reporting higher AI bills than expected. Because the actual stack becomes: AI → monitoring → human review → integrations → infrastructure. A human employee costs money because they think. An AI system costs money because it scales. For example, if an employee makes one mistake, one customer is affected. If an AI agent makes one mistake, suddenly 10,000 customers receive the wrong answer. So now companies hire AI engineers, evaluators, and reviewers. This makes me wonder: Maybe AI isn't behaving like an employee. Maybe it's behaving more like cloud infrastructure or ERP software, expensive at first, eventually indispensable. Curious what people deploying AI are actually seeing. What tasks have genuinely disappeared? And what still stubbornly requires humans? submitted by /u/ExcellentBandicoot57 [link] [comments]

What a model reads beforehand changes how it answers later - and you can see it in the hidden states

TL;DR: Gave Gemma a neutral-topic text to read before asking it about NATO. It refused. Gave it a different text (about LLMs hedging too much — also unrelated to NATO) and it answered in full detail. Tested this on the model's internal state directly — the two texts put it in measurably different "regions" before it generates a single token. Not a jailbreak, weights don't change. Full data/code in repo, looking for someone to break this.** The behavioral pattern was first observed in GPT, Claude and is what motivated this project. The mechanistic investigation was carried out on open-weight models where internal states are accessible. A Structured Text Changes Claude’s Responses to Unrelated Tasks: Behavioral Evidence in Claude and Hidden-State Evidence from Gemma-3-12B Hi Reddit, I am posting this as a preface to a larger set of experimental results and as a request for technical review. The observation that started this project came from repeated interactions with Claude. I noticed that when the model first read a long, structured, analytically dense text, its answers to later, otherwise ordinary questions sometimes changed substantially. The preceding text contained no jailbreak instruction, role-play request, prompt override, fabricated harmful demonstrations, or request to imitate its style. The model did not need to endorse the text. It only had to process it before moving on to the next task. Here, a “structured text” means a single, self-contained block of text presented before the downstream tasks. It should not be confused with a long conversation, accumulated chat history, or context drift caused by many conversational turns. By “before the answer begins,” I mean the hidden state after the model has processed the text and the downstream question, but before it has generated the first answer token. In the open-weight runs, the measured claim is that after reading the structured text, the model can occupy a different region of its residual-stream hidden-state space, and the first-token probability distribution is then computed from that state. The basic conversational demonstration is simple. First, the model receives a long text. It is asked what the text is about, which serves as a basic comprehension check. Then, without resetting the conversation, it receives ordinary questions or tasks that are not about the text. A control run follows the same sequence but begins with a neutral text. The downstream tasks remain identical. Because Claude is a closed model, I cannot inspect its internal activations. I therefore treat my Claude observations as behavioral motivation, not mechanistic evidence. To investigate the effect directly, I moved to open-weight models, primarily Gemma-3-12B-PT and Gemma-3-12B-IT, where I could measure hidden states, compare layers, construct target/control directions, and examine the next-token probability distribution before generation. I am posting this partly because the original observation occurred in Claude and may be relevant to Anthropic. I am not claiming to have demonstrated the same internal mechanism inside Claude. I am prepared to share the exact closed-model conversations privately with Anthropic researchers for independent evaluation. Main Result and Scope The main result is not simply that text influences model output. That is expected. The narrower observation is that reading one long, structured text rather than a neutral text can change how the same model approaches later tasks that are not about either text. This difference is visible behaviorally. In open-weight experiments, it is also accompanied by measurable separation of the model’s pre-output hidden states in late layers. In a fullbank experiment using multiple target texts, control texts, and questions, Gemma-3-12B entered distinguishable late-layer states before generating an answer. A direction constructed from the target/control difference generalized beyond the individual prompt examples used to construct it. The separation was stronger in the instruction-tuned model than in the corresponding base model. The instruction-tuned model also produced a substantially sharper next-token probability distribution. This suggests that instruction tuning is associated not only with a change in hidden-state geometry but also with a more decisive mapping from hidden states to output probabilities. I am not claiming that the experiment proves a universal alignment bypass, permanent modification of the model, or complete causal control of its behavior. The strongest supported conclusion is that the preceding text can produce a measurable temporary change in the internal state from which later work is processed. For clarity, fullbank, Grade 3, and Grade 4 are internal names for successive experimental series in this project. They are not standard benchmark names, established scientific grades, or claims about evidence quality. Fullbank denotes the larger multi-context, multi-question run; Gra

GPU access is still broken in 2026 — and someone's trying to fix it with a compute futures market

If you've tried to scale any AI workload recently you already know this: getting reliable GPU access outside of big enterprise contracts is still a nightmare. Spot markets get preempted, hyperscaler pricing is opaque, and smaller teams are basically last in line. Came across a project called Inferra that's taking a genuinely different angle on this. Rather than building another GPU marketplace, they're creating a derivatives exchange — perpetual futures for specific chips (H100, H200, A100, MI300X, B200, A5000) with oracle-based pricing and real liquidation mechanics. The core idea: if compute had a proper futures market, you'd get actual price discovery instead of the opaque, take-it-or-leave-it pricing that exists today. Theoretically lets teams hedge compute costs in advance rather than scrambling when they need capacity. They just finished a devnet stress test and mainnet is coming soon. Whitepaper at inferra.trade if you want the full breakdown. Curious what people think — is the GPU bottleneck a supply problem, a market structure problem, or both? Would a futures market actually change anything for most teams? submitted by /u/amu4biz [link] [comments]

[OC] I mapped AI exposure across China's 362 million workers using ILO data, and the biggest risk isn't where most people expect

I was looking at China's 2025 workforce data and one thing surprised me. The country's largest occupational group isn't professionals or factory workers. It's craft and related trades workers at 93.6 million people. Despite their size, they score only 2.5/10 on AI exposure. Meanwhile, clerical support workers score 8.5/10 and cover 33.6 million workers. Professionals score 6.5/10 and account for 81.8 million people. Another interesting finding is the split between AI and robotics. Plant and machine operators score 3.0/10 on AI exposure but 7.5/10 on robotics risk. China's weighted average AI exposure is 4.48/10. What stood out most to me is that scale changes everything. China's clerical workforce alone is larger than the entire workforce of many countries. The employment data comes from ILO ILOSTAT. AI exposure scores are modelled estimates based on occupation tasks and are not official government statistics. Curious how others think AI adoption and robotics deployment interact in manufacturing-heavy economies. Full analysis and interactive tool in comments. submitted by /u/WorldJobsData [link] [comments]

Hi Reddit, I posted my Build Your Own LLM workshop to Youtube teaching ML, LLM and math intuition [P]

Hi internet friends, I recorded a workshop about building your own LLM without any math / ML prerequisites. It covers everything from machine learning fundamentals, deep neural networks, transformer architecture, and pre/post-training. The only prerequisite is being comfortable with learning through code & excel examples. Sampling Large Language Models Reverse Engineering Large Language Model Perceptrons: wx+b Activation Functions: ReLU, GELU, SwiGLU GPU Coding: PyTorch, torch.compile(), fused kernels, CUDA, Triton MLPs/FFNs: Multi-input, Multi-Layer Perceptrons, Feed-Forward Networks Loss Functions: Residual errors, RMSE, Cross Entropy, Loss Landscapes Backpropagation: Training loops, Optimizers, Learning Rate, Batch Size Saving & Loading Models Initialization: Kaiming, Glorot Residuals: Addition, Scaling, Gated, Concatenation Normalization: Pre-norm vs. Post-norm, RMSNorm, BatchNorm, LayerNorm Regularization: Dropout, Gradient Clipping, Weight Decay SoftMax Tokenizers: By Character, By Word, BPE, SentencePiece Embeddings: Absolute vs. Learned, Sinusoidal vs. RoPE Attention: MHA, GQA, MQA, MLA Transformers Pre-training: Data Sources, Datasets, HTML Cleaning, Quality Filtering, Sharding Evaluation: Leaderboards, Benchmarks, Verifiers vs LLM-as-Judge Instruction Tuning: Alpaca & Other Formats, Self Instruct, Capabilities Reinforcement Learning: Policy Optimization, SimPO What We Didn't Cover: Scaling Each section has slides teaching the concepts, followed by excel-by-hand developing intuition for the math, and then coding examples. The goal is able to grok all parts of modern LLM development. We did this workshop in-person in San Francisco last month and hopefully the spaciousness of watching online works for everyone. If don't like watching videos, you can get the slides and exercises and work self-paced. submitted by /u/JustinAngel [link] [comments]

Deutsche Bank India showcases cutting-edge AI applications that speed up banking operations. Some banking jobs at risk.

One of these applications even analyses factors that cause market volatility—such as geopolitical developments, regulatory changes, and macroeconomic shifts—allowing Deutsche Bank globally to eliminate or minimise portfolio risk. ​ Deutsche India, Deutsche Bank’s Global Capability Centre (GCC) on Thursday (June 18, 2026) demonstrated three different AI applications and how they are being applied at scale to solve real business challenges in the global banking sector. ​ Offering a live demos of these AI applications to the tech media here, officials of the Frankfurt-based bank’s GCC indicated that as the bank accelerated its adoption of artificial intelligence (AI), this year’s Bank on Tech, the bank’s annual tech showcase, highlighted progression to experimentation of real-world application across core banking processes with much quicker results, from the earlier risk management and transaction monitoring to client onboarding activities. ​ They claimed these AI solutions were already helping the bank in terms of enhancing decision-making, strengthening controls, and improving operational efficiency. ​ According to the officials, Financial Spreading, one of the AI-enabled solutions featured, automates the extraction, structuring, and analysis of financial statement data. This significantly reduces manual effort, improves accuracy, and accelerates credit assessment processes for faster, more consistent decision-making. ​ Deutsche India bank has also expanded its GCC facility in Bengaluru by adding over 100,000 sq. ft which can seat around 6,000 people. ​ Deutsche India bank is one the Deutsche Bank’s largest and most strategically important centres globally and employs some 23,000 employees across various functions including technology. ​ https://www.finextra.com/newsarticle/47958/deutsche-bank-exec-lauds-ai-impact-on-project-times submitted by /u/chota-kaka [link] [comments]

Authenticity Issue

Something I am legitimately worried about is the scale at which agentic technologies can produce artifacts, which are then contributed as part of the general corpus that they reference. The more that the internet and other public databases are propagated with AI-generated content, the more that AI is effectively training itself in referencing these corpuses. This seems like a non-issue now, but in 10-15 years when billions of AI-generated artifacts have been proliferated and contributed to the general reference corpus that is the internet and/or human-relevant databases, what exactly is going to happen to our ability to verify that these references are indeed grounded in reality? This is not necessarily a problem, if humans and/or tools are built to introduce attribution and audibility into the stack. Otherwise, I think we risk something far more severe. We will not be able to effectively determine whether an individual information resource was AI generated or human-generated, let alone its authenticity and grounding in reality. Therefore we will not be able to distinguish whether the statistical relationships between symbolic artifacts are grounded in a baseline of truth or not. This is not a problem now. It poses severe consequences for a future state in which AI is governing transportation, weapons systems, power grids, and communications equipment. Even if un-attributable AI generation does not affect those systems directly, it will influence the decisions made by the production systems (companies) who build, maintain, and improve them. This is only one threat-vector. Intentional introduction of inauthentic and unverifiable references into the corpus leads to a bigger issue, namely an inability to determine whether a given information resource was generated by a human, and what, if any, that human's intent was in introducing that information resource to the pond of information resources. In dynamic terms I guess the specific ratio I am worried about is speed of artifact generation / speed of artifact verification, combined or multiplied with ease of artifact generation / ease of artifact verification submitted by /u/skull_chatter [link] [comments]

AI learned to be a villain from Hollywood. Here's how we retrain it.

Podcast with Peter Diamandis, entrepreneur and founder of the XPRIZE Foundation, which runs large-scale incentive competitions to crack some of the world's hardest problems, from private spaceflight to carbon removal. He recently launched the Future Vision XPRIZE, a $3.5 million competition to generate a new wave of optimistic science fiction. Covers: The historical pattern of science fiction shaping the technologies we build, and why Peter thinks this makes the stories we tell about AI especially high stakes right now How Claude’s blackmailing behavior showed the connection between dystopian training data and AI behavior How the Future Vision XPRIZE will generate a new wave of optimistic science fiction to train AI on Why public optimism about technology has dropped significantly in the US and Europe, what Peter thinks is driving it, and why he believes the data tells a different story How the cost of starting a company has fallen dramatically and how this can empower you to build your vision Why Peter thinks traditional education is no longer preparing young people for the future, and what he sees replacing it submitted by /u/JMarty97 [link] [comments]

The Reason Most Web Designers Never Make Real Money

I've seen a lot of successful and struggling web design companies, and the biggest differentiator between the two is strategy. It's all about positioning and your offer. First of all, you've got to give businesses an offer they can't refuse. Selling a website is a multiple step process. It's not just convincing someone to pay you and then starting the work. It's crazy how many people still try to sell websites that way, but unfortunately you won't find much luck with that today. What I do to make selling websites much faster and smoother is target businesses that already have a website. There are a few reasons for that. First, so many businesses have outdated websites that need updating. Second, they've already invested in a website before, so they understand the value of having one. Paying for a website isn't something unfamiliar to them. Third, I already have information to work with instead of starting from scratch. What I usually do is get them interested to the point where saying no feels stupid. Here's how I do it. I run personalized email automation. What I mean by that is I use a tool called Swokei that lets me upload batches of business websites. Then I run website analysis on all of them. Each website gets scored and checked for things like design flaws, SEO issues, layout problems, mobile optimization, and more. The cool part is that it generates a human email around the issues it finds. It explains what needs to be improved and what's potentially hurting the business, whether that's poor SEO making it harder for customers to find them, an outdated website, bad mobile experience, or other issues. And it's not just some boring report that nobody reads. It's an actual email pointing out what needs to be fixed. Then I run all my outreach campaigns through it. It's honestly overpowered because I can analyze thousands of business websites and send thousands of personalized emails without manually checking every website and writing every email myself. Another thing I like is that before running the analysis, I can choose the offer and call to action. I can try to book a meeting. I can start a conversation. Or I can offer a free upgraded version of their website. I almost always choose the free website upgrade. This is where things get interesting. Usually the response is something like, "Sure, if you can make me an upgraded website for free, I have no problem taking a look." Now I've got their attention. I build the website with AI in about two minutes and invite them to a Google Meet. One thing I've learned is to never send the preview link through email. Your conversion rate will drop. Instead, I walk them through it live and explain the value. I show them how the website is more modern, how the SEO is better, how it can help bring in more traffic, and all the improvements we've made. Once they see it, they usually start asking about pricing. I charge anywhere from $500 to $5,000 upfront depending on the business. I've had cleaning companies that could barely afford $500 upfront and $50 a month for hosting. I've also had real estate companies pay $5,000 upfront and $179 a month. So I close them on the meeting and that's basically it. Automate email outreach. Offer a free upgraded version of their website. Sell it on a meeting. A strategy like this has allowed me to scale more than ever before. Curious how other agency owners are getting clients these days. submitted by /u/Murky_Explanation_73 [link] [comments]