January 9, 2026

When people are substituting mankind with machinekind, there are certain collateral damage happens. It is quite normal, luddites of AI era, have been on the rise. Here what is being claimed, the changes in the structure of brain. New anatomical evidence reveals the shocking difference between AI users and AI purists. According to a rigorous scientific diagram, if you strictly refuse to use LLMs, your brain doesn’t merely remain human. It evolves into a massive, highly complex organ capable of sonar and eating raw fish. You basically become a bottlenose dolphin. 🐬 Some critics kind of imply: “Those who don’t use AI are dolphins: the brightest and smartest creatures swimming among us ChatGPT zombies.” Funny line. Weak model. Because the real dividing line isn’t AI vs no AI. It’s passive use vs active use. Let me show you the anatomy of the Nueral Machine ( Machine Brain). And Credits: Graphcore. Apply for 500+ Semiconductor jobs in Graphcore, Bangalore. Graphcore, now a wholly owned subsidiary of SoftBank Group, will invest up to £1 billion over a decade to establish an AI Engineering Campus in Bengaluru. The plan targets ~500 semiconductor jobs, with ~100 roles hired immediately across silicon logic/physical design, verification, characterization, and bring-up. The expansion in India is paired with a UK headcount increase to ~750, primarily in silicon, software, and AI engineering. The campus will contribute to Graphcore’s end-to-end AI compute stack (silicon, systems, software). SoftBank’s broader AI push includes large-scale infrastructure initiatives such as Stargate, in partnership with OpenAI and Oracle. SoftBank reports >$12 billion invested in India over the past decade. Rationale (explicitly stated): Bengaluru’s concentration of universities, startups, and multinational tech talent. Alignment with Indian government semiconductor initiatives and skills development. Implications (inference, bounded): Signals deepening global semiconductor R&D presence in India, not just services. Positions Graphcore to scale AI silicon development capacity while diversifying talent geography. Reinforces SoftBank’s strategy to vertically integrate AI infrastructure—from chips to platforms. Bottom line: This is a long-horizon R&D and talent bet, anchoring advanced semiconductor engineering in India while integrating Graphcore into SoftBank’s broader AI infrastructure agenda. Career at Graphcore Director / Senior Principal AI SoC Validation (Bring-up lead)Engineering – SiliconApply now Principal Bring-Up and Characterisation EngineerEngineering – SiliconApply now Principal Embedded SW/FW Engineer (Bringup) – Bengaluru, multiple vacanciesEngineering – SiliconApply now Principal Silicon Physical Design Engineer – BengaluruEngineering – SiliconApply now Principal Silicon Verification Engineer – BengaluruEngineering – SiliconApply now Senior Bring-Up and Characterisation Engineer – BengaluruEngineering – SiliconApply now Senior Bring-Up and Characterisation Engineer – Bengaluru, Multiple VacanciesEngineering – SiliconApply now Senior Bring-Up and Characterisation Engineer – Bengaluru, Multiple VacanciesEngineering – SiliconApply now Senior Embedded SW/FW Engineer (Bringup) – Bengaluru, multiple vacanciesEngineering – SiliconApply now Senior Embedded SW/FW Engineer (Bringup), Bengaluru, multiple vacanciesEngineering – SiliconApply now Senior Silicon Physical Design Engineer – BengaluruEngineering – SiliconApply now Senior Silicon Verification EngineerEngineering – SiliconApply now Silicon Physical Design Engineer – Bengaluru, Multiple VacanciesEngineering – SiliconApply now Staff Bring-Up and Characterisation EngineerEngineering – SiliconApply now Staff Bring-Up and Characterisation Engineer – BengaluruEngineering – SiliconApply now Staff Embedded SW/FW Engineer (Bringup) – BengaluruEngineering – SiliconApply now Staff Silicon Logical Design Engineer – Bengaluru, Multiple VacanciesEngineering – SiliconApply now Staff Silicon Physical Design Engineer – BengaluruEngineering – SiliconApply now Staff Silicon Verification Engineer – Bengaluru, Multiple VacanciesEngineering – SiliconApply now The Connotation We’re Beating: “Using AI Makes You Passive” There’s truth inside the fear. If you stop exercising a muscle, it atrophies. If you use AI to replace thinking, you’ll get worse at thinking. That’s what “getting dumber” looks like in practice: blindly copy-pasting code you can’t explain, auto-generating strategy you can’t defend, letting autocomplete decide your business logic, accepting output without checking assumptions. But here’s the distinction people keep skipping: Outsourcing thought and outsourcing execution are two different things. AI can accelerate the HOW—drafting, refactoring, summarizing, exploring options. Humans must own the WHAT—what matters, what to build, what tradeoffs to accept, what “good” looks like, what you’re willing to be accountable for. Passive consumption makes you dumber. Active use makes you sharper. So how do you force active use every time? That’s where Sanjaya Uvacca comes in. Sanjaya Uvacca: Not “Just a Narrator,” But a Discipline of Clarity “Sanjaya uvāca” literally reads as “Sanjaya said…”—and it’s easy to treat that like a speaker label you skip past. But Sanjaya isn’t a decorative narrator. He’s an archetype: the lucid witness who can stand near chaos without being swallowed by it, and transmit what matters without distortion. And that gives us the cleanest metaphor for using AI well: In the Mahabharata, Sanjaya is the bridge between battle and blindness: he translates an overwhelming battlefield into clear perception for someone who cannot see it directly—without taking the choice away from him. That’s the most accurate metaphor for Artificial Intelligence. Our “battle” is complexity (information overload, shifting contexts, endless variables). Our “blindness” is cognitive (limited attention, limited time, partial knowledge). Used well, AI becomes a Sanjaya: it expands perception—summarizes, compares, surfaces risks, reveals tradeoffs—so the human can answer the harder question of what matters and what to do. Used poorly, AI stops being a bridge and becomes a crutch: it doesn’t cure blindness, it replaces judgment. So the goal is not “use AI” or “don’t use AI.” The goal is: use AI like Sanjaya. Perception amplifier. Clarity partner. Reality translator. Not a substitute decision-maker. The Recursive Window: Features That Make Passive Use Hard I don’t rely on a “magic prompt.” I rely on a workflow—a recursive window that forces thinking before output. Here are the features: 1) Start with purpose and constraints Before generating anything, the process forces you to define: what outcome you want, who it’s for, what constraints matter (time, budget, risk, tone, scope). This immediately locks you into the human job: WHAT are we trying to do? 2) Build a map before building the artifact Instead of jumping to a final answer, you first create structure: a syllabus for learning, a plan for execution, dependencies and “unknowns” made visible. This prevents the

AI TRUST FRAMEWORK (ATF) By lakshminarasimhan santhanam Iyengar,  Mentor@QAF Lab India As AI systems generate ever more fluent text, “AI slop” has emerged as a term for output that reads well but is low-quality, unverified, or misleading. Such slop content – generic, citation-free text filled with surface-level claims or even hallucinated details – can erode user trust. To increase human trust in AI systems, developers are proposing metrics that make slop observable, reduce the human burden of fact-checking, and support better trust calibration through transparency and confidence signaling. Below, we survey key metrics (grounded in recent research and industry practice) aligned to an “Anti-Slop” framework, covering: (1) measuring slop, (2) human verification costs, and (3) output trustworthiness/transparency. 1. Measuring “Slop” in AI Outputs AI slop generally refers to a flood of generic, unverified AI-generated content that prioritizes quantity over quality. To systematically identify or quantify “sloppy” outputs, researchers have defined interpretable metrics that flag telltale signs of low-quality or fabricated content: Information/Claim Density: Sloppy text often packs in numerous factual-sounding claims without substance. Factual Claim Density (FCD) is one proposed metric counting how many statements appear to assert verifiable facts per 100 words. A higher claim density can indicate more opportunities for error, making slop easier to spot. Example: An answer that rattles off many specific “facts” (dates, statistics, quotes) has high FCD – if these aren’t all validated, trust should drop. Conversely, lowering FCD (fewer unsupported facts) can mitigate hallucinations. In a similar vein, information density measures the ratio of meaningful content to filler: truly sloppy passages tend to be verbose yet say little of value. Researchers compute this via token entropy or idea density – low information density (lots of words but little new information) is a strong slop indicator. Redundancy & Genericness: AI slop often features repetition and template-like prose (e.g. the same stock phrases or generic statements that “could apply in almost any context”). Metrics like lexical diversity or compression ratio can capture how repetitive an output is. High redundancy signals low originality, often correlating with content that was cheaply auto-generated. Generic answers that stay vague and avoid specifics can also be flagged – they may feel coherent but dodge providing verifiable detail, another slop tactic. For instance, Shaib et al. (2025) note that verbose answers full of broad, non-specific statements are a hallmark of slop. Measuring the proportion of vague or boilerplate sentences thus helps quantify sloppiness. Unverifiable Claims & Hallucinations: Perhaps the clearest sign of slop is factual unverifiability – claims that cannot be confirmed by any reliable source. Metrics here include hallucination rate, defined as the percentage of outputs containing at least one unsupported claim. Automated pipelines now use retrieval or fact-checking models to assess each answer: e.g. a “claim grounding check” can be done by re-querying a verifier model on each sentence to ask if it’s supported by provided sources. An elevated hallucination rate or a high unsupported claim density (unsupported claims per 100 answers) directly quantifies the slop in a system’s output. Failure mode: A model might produce a very detailed-sounding answer that is entirely invented; by counting how many such unsupported details slip through, teams can track slop reduction efforts over time (with a goal of driving the hallucination rate down, e.g. <5% of answers with any unsupported claim). Citation Quality and Support: Requiring AI to show its sources makes slop far more observable. Sloppy content is typically “citation-free” or uses irrelevant references, whereas trustworthy output should link each significant claim to a verifiable source. New metrics from retrieval-augmented generation (RAG) evaluations explicitly score citation precision (what fraction of cited sources truly support the claims) and citation coverage (what fraction of the answer’s statements are backed by the provided citations). High citation precision and coverage mean the model’s claims are well-supported by evidence on average, boosting trust; low scores reveal that many claims are unsourced or improperly sourced (a red flag). In practice, source-attribution checks are being built into editorial pipelines – for example, requiring every factual claim to include a footnote or URL and then auditing whether those links actually verify the text. Enforcing this not only catches hallucinations but deters them: “unsupported claims become visible and auditable” when every statement is expected to map to a citation. This makes it much harder for slop to hide. Specificity Risk Flags: While slop is often overly generic, another failure mode is spurious over-specificity – i.e. the model confidently adds extra details not in source material. For example, a hallucinated answer might insert a precise statistic or a proper name to sound convincing (“In 2022, 47.3% of X…”), even though no such detail was provided. Some frameworks propose flagging high-specificity content for verification, on the premise that the more specific a claim, the easier it is to check – or to catch if false. One approach is to count the number of entities, dates or numeric figures in an answer as a “specificity score”; unusually specific answers, especially absent citations, carry higher factuality risk. In essence, when an LLM goes beyond the given information to provide granular facts, that “extra specificity” is a risk signal that might warrant an automatic caution or human review. While not yet a standardized metric in academic literature, this idea aligns with check-worthy claim detection research – identifying which statements “appear to assert something factual” and thus merit fact-checking. In practice, pairing specificity flags with retrieval can mitigate slop: if the model gives a highly detailed answer, the system can be tuned to either provide a source for each detail or explicitly abstain from unsupported specifics. Use cases: These slop-focused metrics are being tested in both research settings and production. For instance, Shaib et al. (2025) developed a “slop taxonomy” and showed that human judgments of sloppiness correlate with factors like low information density, irrelevance, and factual errors. On the industry side, developers of knowledge assistants track unsupported claim rates and tune models to reduce them (e.g. via retrieval or by penalizing high FCD in