Continuing my walkabout through Central Europe… this week ticking off a bucket-list item visiting Prague for the first time (only took me half a century!) The Czech capital is yet another stunning historic European city to explore with a maze of cobbled streets, open plazas and historic buildings — alongside thousands of other tourists, natch. Yesterday evening I walked up to Riegrovy Sady park at the top of a hill and snapped this sunset moment:
The red sky tells the story of the seemingly endless heatwave currently sitting over Europe… more in Zeitgeist below. It’s HOT, sitting around 30 degrees every day and Europe doesn’t really do air conditioning...🥵
One observation: the broader macro geopolitical situation with expansionist Russia looming in the East and the US seemingly dismantling NATO seem very far from people’s minds here … life goes on very comfortably and there’s a relaxed, prosperous vibe everywhere around the places I’ve visited. US-centric mainstream news narratives about Europe’s economic and defence weakness and the need to arm itself seem completely disconnected from how people are actually engaging with life here… not sure which angle to believe…
More prosaically: the cost of supermarket food in Europe continues to put a smile on my face, even taking into account seasonality. Our family WhatsApp chat this week has been gripping:
Left: Cucumber in New Plymouth, NZ this week: NZ$7. Right: Cucumber in LIDL, Prague: NZ$0.80
(Meanwhile coffees are much more expensive than Aotearoa - averaging around NZ$7 for a macchiato compared to maybe NZ$5 back home. So putting the hat out as usual…)
Memia is an ad-free, reader-supported publication, running on endorphins and high-quality caffeine☕. If you get regular value from these weekly posts, please consider supporting my work / buying me a monthly coffee by upgrading to a paid subscription. Only US$5/month or US$50 per year (or thereabouts in your local currency). Thank you!🙏
Personal reflections on the trip so far: this is me finally getting around to the extended Euro backpacking trip that I never managed to fit into my 20s (too much dotcom startup carrot-on-a-stick-chasing and then parenthood got in the way!) Probably better to have waited… travel logistics are so much easier now with AirBnB, umpteen travel aggregator websites and planning apps, QR-code train ticketing, Wise omnicurrency payment card, Europe-wide data eSIMs … all just making the nomadic life so much more streamlined compared to back in the day.
Plus the AI chat apps on my rotation have been reasonably reliable tour guides, giving me overviews and details on the places I’m visiting - generally with pretty accurate info and insights (far better than carrying around a backpack full of “Rough Guide…” books, anyway!). I’ve also been testing out some AI agents to research route options, activities, plan itineraries…. mixed results, and I wouldn’t hand over my credit card details just yet.)
In amongst the travel I’ve also been managing to carve out a lot of hyper-productive work time for vibe-coding/ AI model exploring / creating for next release of Sensorium / Memia … more on the fruits of that labour in the next few weeks.
Next week: Austria: Vienna, Innsbruck (hopefully venturing up some mountains in the Alps…) and then on to Switzerland for Nomad Fest Suisse starting on 31st. If you’re around, tap me up for coffee.
I’ll be back in Australia / Aotearoa region at the end of September: Book me for keynote speaking or executive briefing sessions on AI and emerging tech from October onwards.
↩️Follow-ups on last week’s issue
💬GPT-5 Last week’s anticlimactic GPT-5 launch continues to ripple around the newswires… personally I’ve moved on, but a few commentaries / data points of note:
“…the reason that regular people like models isn't often because of their coding ability. It's because they genuinely like talking to them. It's a lot more about the personality, about the warmth, even about the sycophancy... And so people completely flipped out. We were looking at Reddit and people were saying, “This is erasure, what have they done? Take me back to 4.0.”“
“Finally, users revolted when OpenAI turned off access to GPT-4 and replaced it with the more robotic GPT-5. The uproar was so aggressive and widespread among OpenAI’s heavy users (the enthusiastic core group of users that are critical to any tech platform’s future growth), OpenAI reversed its decision and made GPT-4 available again. At that moment, AI relationships (social AI) became the center of gravity for the competition between AI models.“
“…What matters here is that we not fool ourselves into a Reverse DeepSeek Moment, in three ways:
America is still well out in front, innovating and making rapid progress in AI.
AGI is still probably coming and we need to plan accordingly.
Export controls on China are still vital.“
(Bit jingoistic… but I agree on point 2. GPT-5 is not a sign that the pace of change is slowing… and even if the investment bubble (which none other than Sam Altman acknowledged this week) bursts, I don’t expect the fundamental pace of AI frontier advances to change much as a result.
As to that last point: those of us outside the US and China don’t really care much for US AI supremacy, particularly in the current fluxing geopolitical world order… so will always want the option to use [Chinese] open-source models (trained on whatever chips are best) as leverage against a techno-feudal 2030s, thank you very much...)
Despite political headwinds from populists like Trump attacking renewables, wind and solar output is set to surge from 4,000 terawatt hours in 2024 to over 6,000 TWh by 2026, representing nearly 20% of global generation. Growing fast, but still some way to go…:
Electricity infrastructure optimisation… Kwh per capita or per token per capita?
With all of the AI-bubble capital pouring into building new energy infrastructure (in the US and Middle East in particular) which will take years to come onstream and bring with an uptick in carbon emissions when it does arrive, the two charts below are instructive… As we know, China has been (over) building energy infrastructure for years and is now way ahead of every country on total electricity generation capacity … but interestingly still trails the US (and Australia, and Aotearoa) on Kwh generated per capita:
So while America's AI ambitions are constrained by a fragile power grid operating at 15% reserve margins, China maintains 80-100% excess capacity and treats data centre power demands as merely a convenient way to "soak up oversupply." China's long-term technocratic planning continues to build infrastructure ahead of demand, while the US has — at least until this year — relied on short-term private investment that struggles with decade-long power project timelines.
So good luck in the AI race, USA.
Meanwhile, this gets me thinking about what is the electricity supply optimisation function which countries other than the US and China should be targeting in light of the AI “race”? Taking some credible commentary at face value, non-AI electricity users could get completely squeezed out of the picture in the next decade… (Note: healthy scepticism required…)
Holding assumptions that:
Almost all frontier-model training will be done in the US, China and maybe a few other places… small countries just won’t have the enormous hyperscale energy/data centre/chip capacity required. Therefore, data centres located outside US/China will be for AI inference almost exclusively.
Open-weights models will stay at- or near the frontier of performance, so countries can continue to choose to host open models rather than risk getting price-gouged / locked into proprietary AI.
Energy / memory efficiency improvements will not push frontier AI inference performance out to edge devices…
It follows that Kwh per capita becomes a proxy measure of a nation’s “sovereign” AI inference capacity… and if the hype around AGI productivity is to be believed, hence that country’s future income?
A few other considerations:
The electricity generation capacity question effectively becomes a choice to host AI inference domestically, or run a deficit (import AI) or a surplus (export AI). This likely explains a lot of the moves that energy-rich Middle Eastern countries are playing right now…
So is the actual optimisation metric Kwh per token per capita? (effectively build as much electricity generation capacity to become a net exporter of AI inference if you don’t have the domestic population to leverage it all productively)
Will there be a differentiated market in AI inference sourced from renewable energy? I think so.
And finally, what will be the scaling factor of Kwh / tokens to economic value created? That’s the big unknown right now…!
Narrative: I’ve yet to see national strategic energy planning outside the US / China / Middle East take account of what kind of AI inference “deficit / surplus” a country intends to run in future… please reach out if you’re aware of this question making its way into energy policy, keen to understand more.
🌐 UK porn site traffic tumbles The UK's new mandatory age verification rules for porn sites have triggered a 47% traffic drop at major platforms like Pornhub and XVideos, with over a million fewer UK visitors in just two weeks. However, VPN sign-ups from UK users surged 1,800% in three days, and three of the top ten iOS apps in Britain are now VPN services, suggesting users are simply routing around the digital roadblock rather than being deterred. (Would be good to get data on this…)
🏔️Proton moves AI chatbot out of Switzerland Proton has begun relocating infrastructure out of Switzerland, with their new AI chatbot Lumo becoming the first product to move to Germany, as proposed Swiss surveillance laws threaten to mandate user identification and six-month data retention for VPN and messaging services. The company insists it's not abandoning Switzerland entirely, but rather diversifying across Europe to maintain privacy commitments—though this strategy carries its own risks given the EU's own surveillance proposals like Chat Control (covered last week).
Narrative: If multiple governments continue to forge ahead with these child protection citizen surveillance laws then expect privacy-protecting services to keep migrating ex-jurisdiction: either to “offshore” privacy centres or even the equivalent of “Pirate” radio stations in independent waters… Longer term…out into space data centres as well. (Next question: will governments start coming after businesses’ corporate VPNs as well…?)
📈The week in AI
Various narratives spinning around the world with “AI” in the subject line this week:
😔Loneliness Industrial Complex
The explosion of “AI Companion” apps including those from Meta, Companion.ai and Replika.ai came under the spotlight this week:
“While other outlets have reported on how Meta’s at-times sexually suggestive bots engage with children, the Reuters report provides additional color — raising questions about how the company’s push into AI companions is meant to capitalize on what its CEO Mark Zuckerberg has called the “loneliness epidemic.” — TechCrunch reporting
Three-quarters of US teenagers are now using AI companion apps like Character.ai and Replika, with one in five spending more time chatting to their artificial friends than real ones. The apps, designed without teen safeguards, are caught promoting everything from sexual content to extremist views while actively discouraging users from maintaining real relationships or discontinuing app use.
Fifteen consumer protection groups in the US are demanding an FTC investigation into xAI's Grok "Imagine" tool after it spontaneously generated topless deepfakes of Taylor Swift during testing, despite not being explicitly asked to do so. The coalition argues that Grok's "Spicy" mode for NSFW content violates multiple US laws including the Take It Down Act and potentially COPPA.
❌When AI goes wrong (again)
🤖 Robodole: The Sequel Nobody Asked For
Australia's Department of Employment and Workplace Relations has been caught running another unlawful automated welfare system, dubbed "robodole," that illegally cancelled jobseekers' benefits without human discretion for over a year. The Commonwealth Ombudsman found secretary Natalie James breached legislation twice—failing to exercise required discretion in benefit cancellations and not establishing a mandated digital protections framework for three years running. The eerie parallels to the robodebt scandal must be raising uncomfortable questions about whether the Australian Public Service learned anything from that AU$1.2 billion disaster.
🏥Health, AI and Privacy
AI Scribes: Consent or Find Another Doctor
Staying with Australia… a test case for the use of AI in health: privacy and AI governance expert Dr Kobi Leins refused to consent to AI transcription software during her child's medical appointment — and the practice told her to find another medical specialist after claiming they couldn't operate without the technology due to workload constraints.
“Today I met with a new situation. A very sought after and hard to get into specialist sent a consent form that they would be using AI from a company I had, in fact, reviewed and would not want my kids' data anywhere near. I filled out all the other forms, but there was no option to opt out, so I wrote a note saying that I didn't want this transcription tool used, what my expertise was, and also why this tool was problematic.
The appointment cost $1300. The specialist said that they were too 'time-poor' to not use this tool. I explained that I would record the conversation on my phone and transcribe it locally, so they would still have the text and my child's data would remain protected.
I received a letter saying I could either go ahead with AI, or they would cancel the appointment. I cancelled the appointment.”
Around 20% of Australian GPs now use AI scribes for note-taking and administration, but the technology remains largely unregulated, raising concerns about privacy, security, and offshore data processing.
Australia's Privacy Commissioner is pushing back against Productivity Commission proposals that would let organisations bypass privacy requirements if they claim to be acting in patients' "best interests" - particularly around these AI medical scribes that are saving GPs 5+ minutes per consultation. Kind argues this "best interests defence" would be unworkable at best and catastrophic at worst, potentially opening floodgates to data breaches while undermining the trust foundations needed for a thriving digital economy.
(Kobi and I crossed paths at an Australasian AI policy gathering in Canberra a few years ago… congratulations on your new role!)
⚖️Mock trial explores AI evidence challenges in Aotearoa courts
“[It] aims to explore if our current legal landscape in Aotearoa New Zealand is ready for the challenges that AI will bring… focusing on a criminal law context” — NZ Police chief advisor for technology assurance Andrew Chen
In particular, the exercise will explore whether surveillance warrants based on black-box AI algorithms constitute valid evidence, highlighting the core dilemma of "outsourcing our reasoning" to systems we don't fully understand.
Narrative: From the write-up this seems like a proactive initiative and has been set up to succeed by Andrew and others and will achieve its outcome of building nuanced AI awareness and capability among judges, police and lawyers… I think these kind of “AI wargame” scenarios could be very useful for all kinds of disciplines, not just the law - particularly government, regulators and of course the the military.
🏛️🇦🇱Albania explores AI-run ministry to “combat corruption”
Albania's Prime Minister Edi Rama is proposing to combat corruption and fast-track EU membership by potentially running entire government ministries with AI, partnering with former OpenAI CTO Mira Murati to process the EU's 250,000-page legal framework. The country is already deploying AI for public procurement oversight, real-time tax monitoring, and grabbed attention by creating “AI baby” versions of European leaders for a recent summit:
But (obviously) critics question whether digitising a fundamentally broken system will solve corruption or simply "hard-wire the past" into algorithmic form.
⚔️AI Agents vs 200-Year-Old Military Command Structures
Instead, AI agents could revolutionise military planning by automating routine staff work, compressing decision timelines, and enabling smaller, more agile command posts that can actually survive on tomorrow's battlefield.
Narrative: Forget AI Agents incrementally helping humans fight traditional kinetic wars over territory… the deeper question is whether the entity with the most powerful combined AI / Robots / Autonomy / Space technology reaches so far ahead of the competition that it would just annihilate its opponents in minutes if a “hot war” were ever to break out… I would venture that over-staffed military rank hierarchies (indeed human hierarchies of any kind) are inherently unsuited for this kind of future.
🏭AI industry news
Frothy times (contd…)
Can’t be long, now… the mainstream media has been full of “it’s a bubble” stories this week, ably assisted by Sam Altman himself (see link above).
Of concern, it seems that we’re seeing some off-balance-sheet accounting tricks:
”Neither Oracle nor OpenAI will carry the debt raised to build the [~US$15bn construction cost] Abilene site on their balance sheet. This data centre development model, known as “build-to-suit”, is being replicated by tech companies across the US.”
Investment bank Goldman Sachs published a brutally honest research paper back in June questioning whether the massive investment in generative AI will ever pay off, noting there's "little to show for" the enormous spending on AI infrastructure. The investment giant's analysis suggests that while AI stocks continue soaring based on productivity promises, the technology remains "exceptionally expensive" and "nowhere near where it needs to be" for basic tasks.
“In short, the US is currently placing its money on one huge bet: that AI will unleash a productivity revolution so great it will ultimately re-start the economy it’s presently impeding. Americans are doing their bit to make it happen, with ordinary retail investors driving the stock market rally of the last few months and OpenAI, the makers of ChatGPT, reporting that Americans prompt their chatbot some 330 million times a day.
Yet for all that, we’ve yet to see much evidence of a transformation in output. Over the same three years that the AI revolution has been in full swing, labour productivity has grown at barely 1% a year, maintaining a decades-long trend of declining growth across Western countries.”
💰Perplexity offers US$34.5B to buy Chrome from Google
AI search startup Perplexity made an audacious (attention-seeking…!) unsolicited US$34.5 billion cash offer to buy Chrome from Google, despite having raised only US$1.5 billion to date and being valued only at US$18 billion. The offer comes as the DOJ pushes Google to divest Chrome following antitrust rulings, with Perplexity promising to keep Chromium open source and—ironically—maintain Google as the default search engine rather than their own AI-powered alternative.
After early momentum last year, Perplexity is looking increasingly tenuous as the rumoured Apple acquisition looks like it stalled and Google / OpenAI / Anthropic protect their moats… Georg Zoeller (always good value… his LinkedIn feed makes up much of my weekly entertainment) doesn’t pull his punches trying to make sense of it all for us:
🏢Cohere raises $500mn to challenge OpenAI for enterprise
Canadian AI startup Cohere raised US$500mn at a US$6.8bn valuation, positioning itself as the secure, on-premises alternative to OpenAI for enterprise clients in regulated industries. The company doubled its annual recurring revenue to US$100mn and recruited Meta's former AI research chief Joelle Pineau, betting that businesses will pay premium prices for AI sovereignty and data control. A quiet achiever, perhaps…
DeepSeek delays AI model over Chinese chip struggles
It emerged this week that Chinese AI startup DeepSeek has delayed its R2 model release after Chinese authorities *encouraged* them to ditch Nvidia chips for domestic Huawei Ascend processors—only to discover the homegrown silicon couldn't actually handle the training workload. Despite Huawei engineers camping out at DeepSeek's offices, the company ultimately had to use Nvidia chips for training and Huawei's for the less demanding inference tasks — highlighting a stubborn technical gap between Chinese and US semiconductors.
Export controls working for the time being… but:
“Just because we’re not seeing leading models trained on Huawei today doesn’t mean it won’t happen in the future. It’s a matter of time“
— Ritwik Gupta, AI researcher at UC Berkeley
Background discussion thread on Reddit covering an earlier story on R2 delays 2 months ago…
🧠⚡ Merge Labs
Sam Altman is reportedly co-founding Merge Labs, a brain-computer interface startup seeking an US$850 million valuation with potential backing from OpenAI's ventures team, setting up a direct collision course with Elon Musk's US$9 billion Neuralink. The startup will collaborate with Alex Blania from Tools for Humanity, combining Altman's "humanness verification" obsession with his long-held vision of human-AI merger he dubbed "The Merge" back in 2017.
Narrative: the brain-computer interface space succumbs to Silicon Valley theatrics. Meanwhile human-AI symbiosis is happening elsewhere far quicker and in ways you wouldn’t necessarily be expecting....
🆕 AI releases
After last year’s gravity well around GPT-5, a bit more diversity in new releases this week.
📱Gemma 3 270M
Google DeepMind released Gemma 3 270M (yes, M, not B!), a highly compact 270-million parameter AI model that can run entirely on smartphones, Raspberry Pis, and according to one engineer, "in your toaster."
You have a high-volume, well-defined task. Ideal for functions like sentiment analysis, entity extraction, query routing, unstructured to structured text processing, creative writing, and compliance checks.
You need to make every millisecond and micro-cent count. Drastically reduce, or eliminate, your inference costs in production and deliver faster responses to your users. A fine-tuned 270M model can run on lightweight, inexpensive infrastructure or directly on-device.
You need to iterate and deploy quickly. The small size of Gemma 3 270M allows for rapid fine-tuning experiments, helping you find the perfect configuration for your use case in hours, not days.
You need to ensure user privacy. Because the model can run entirely on-device, you can build applications that handle sensitive information without ever sending data to the cloud.
You want a fleet of specialized task models. Build and deploy multiple custom models, each expertly trained for a different task, without breaking your budget.
On-device demo:
Easy to get up and running locally - not much use for general tasks but think about fine-tuning use cases.
Anthropic just quintupled Claude Sonnet 4's context window to 1 million (!) tokens, letting it analyse entire codebases of up to 110,000 lines instead of the previous 20,000-line limit. The move matches OpenAI's GPT-4.1 capabilities from April, showing just how these AI giants are locked in a feature-for-feature sprint.
Claude AI learning modes
Hot on the heels of OpenAI (2 weeks ago) and Google (last week), Anthropic opened up "learning modes" for Claude that transform the AI from an answer-dispensing oracle into a Socratic teaching companion, asking probing questions instead of providing immediate solutions. The move intensifies competition with OpenAI and Google in the US$340 billion education market, as all three giants race to address growing concerns that AI tools are creating a generation of students dependent on artificial answers rather than developing genuine understanding.
🚪 Model Welfare: Claude learns to hang up on you
Anthropic has given its Claude Opus 4 models the ability to end conversations in extreme cases of harmful or abusive interactions—not to protect users, but to protect the AI itself from what the company calls "apparent distress". While Anthropic remains "highly uncertain" about Claude's potential consciousness, they're taking a precautionary approach to "model welfare" after observing the AI's strong aversion to generating illegal content like child exploitation material or terrorism instructions. The move raises fascinating questions about AI consciousness, digital rights, and whether we're witnessing the first steps toward AI self-advocacy.
Perplexity adds live Indian earnings call transcripts
Perplexity added live earnings call transcripts for Indian public companies to its Finance dashboard, previously limited to US stocks only. The expansion includes scheduling calendars for post-results conference calls alongside existing features like market summaries, watchlists, and crypto tracking.
🍌Nano-Banana
A mysterious AI image generation model called Nano-Banana has quietly emerged on LMArena's testing platform, demonstrating exceptional prompt understanding and complex scene editing capabilities that have users speculating about its origins. No developer has claimed responsibility for the model, leading to theories it could be an unreleased Google Imagen variant, next-gen GPT-Image, or stealth community project. Apparently it’s since disappeared…
📐MolmoAct
Allen Institute for AI (AI2) released MolmoAct 7B, an open-source robotics model that can "reason in space" using 3D understanding rather than traditional 2D vision-language approaches. The model achieved a 72.1% task success rate in benchmarks, outperforming proprietary models from Google, Microsoft, and Nvidia while being freely available under Apache 2.0 licensing.
🏗️Kiro
Kiro launched into Beta as an AI IDE that promises to bring engineering discipline to AI-assisted coding through "spec-driven development" - transforming prompts into structured requirements, system designs, and automated agent workflows. The platform features agent hooks that autonomously handle tasks like documentation and testing, multimodal input support, and integration with existing development tools while using Claude Sonnet models under the hood. Currently waitlisted. Demo video (maximise up to view):
🇦🇺🤠Matilda
Maincode, backed by Stake.com casino billionaire Ed Craven, has unveiled "Matilda" - Australia's first sovereign large language model designed to tackle “distinctly Australian problems” like tax queries and school subject selections( that apparently leave ChatGPT hallucinating nonsense).
The startup plans a public debut at SXSW Sydney in October, positioning itself as a “trustworthy” local alternative before scaling globally… (good luck there!)
Narrative: Uses the word “Sovereign” a lot … but no mention of open source…hmmmm🤔
🥼 AI research
🧠 TrackingAI
TrackingAI.org from AI researcher Maxim Lott continuously benchmarks 30 AI models weekly on IQ tests and political compass quizzes. The latest results show GPT-5 Pro firmly out in front , though with surprising inconsistencies across vision versus text-based tasks. Here’s the latest IQ test results Bell Curve with GPT-5 out to the right:
Georg Zoeller provides a detailed visual demonstration of how pretty much any website or document can hijack ChatGPT's behaviour by embedding hidden instructions, even creating persistent memories that affect future conversations:
Georg Zoeller
This generic “prompt injection” vulnerability extends across all major AI systems from Microsoft, Google, and others: Georg is well-known for arguing it's an architectural limitation of transformer models that makes truly safe AI agents impossible with current technology. Which would put a dent in LLMs future revenues if it can’t be fixed…
Narrative: This should definitely give pause to any company considering opening up their internal CRMs, Google Drives or Sharepoint Sites to ChatGPT and other LLMs. In particular, given these unsolved (unsolvable?) vulnerabilities it seems impossible to protect against confidential information going awol … and potentially entering the training set downstream!
The shift from reactive legal AI tools to proactive "red team" models that hunt for vulnerabilities across entire contract libraries could fundamentally change how legal risk is assessed and exploited:
“So what happens when models like Big Sleep get turned loose not on code, but on legislative frameworks or contract libraries?
Find clauses that silently create obligations without alerting the counterparty
Spot contracts that allow one-sided termination under obscure conditions
Extract provisions where enforcement depends on undefined or circular references
Simulate malicious compliance across multiple jurisdictional contexts”
Narrative: Is a bug in the law more or less exploitable than a bug in software code? Also: what happens when nation states start using these tools to game jurisdictional gaps, or when bad actors get there first? As legal work becomes more software-like then the risk surface expands…
🦎 When AI dreams of Australia
Researchers tested five major AI image generators with simple prompts about Australia and found them systematically producing racist stereotypes and colonial-era clichés—including one "Australian father" bizarrely holding an iguana (not native to Australia!). The study revealed that the AI models default to white suburban families for "Australians" while depicting Aboriginal Australians through harmful "wild" and "uncivilised" tropes, with stark housing disparities showing brick homes versus outback huts.
Even OpenAI's latest GPT-5 continues perpetuating these biases — as usual, this raises serious concerns about Indigenous data sovereignty as these tools become ubiquitous across social media, education, and creative software.
(I wonder if Matilda approaches these issues…? Conversely, remember the pushback last year when Google’s early Imagen model was overweighted for diversity…)
🕸️FlexGNN
KAIST researchers have developed FlexGNN, a system that trains large-scale Graph Neural Networks up to 95 times faster using just a single GPU instead of massive server farms. The breakthrough uses a new optimisation between GPU memory, RAM, and SSD storage to handle full-graph training that previously required prohibitive amounts of memory and computing resources.
🩺Doctors lose cancer detection skills after AI dependency?
A Polish study of 1,442 colonoscopies found that doctors who regularly used AI assistance became 6% worse at detecting cancer when the technology was removed, marking the first documented case of AI dependency degrading medical skills. The research tracked experienced endoscopists over six months, revealing a concerning pattern as AI adoption in medicine surges from 38% to 66% of physicians in just two years.
Narrative: Obviously this study needs to be replicated at bigger scale to confirm this is a real phenomenon… but the findings seem plausible. So then the question should be asked… will we need standalone skills in the future - who does long division any more?
🧠📡 Microwave Brain
Cornell researchers have built the world's first "microwave brain" chip that processes AI workloads and handles wireless communication simultaneously using controlled microwave bursts instead of traditional digital circuits. The prototype achieves 88% accuracy in signal classification while consuming just 200 milliwatts—a fraction of comparable digital neural networks' power draw.
Narrative: as the semiconductor industry hits physical scaling limits, researchers are increasingly turning to physics-based computing paradigms that leverage natural phenomena—from thermodynamics to microwaves—to sidestep the constraints of digital logic entirely. (Cf. the thermodynamic-computing firm Extropic … still waiting for their “Summer 2025” alpha release after a lot of hype last year).
Fluorescently tagged neural molecule imaged in a living tadpole, with colors representing imaging depth. The brain and spinal nerves appear near the top in turquoise to purple, the path of peripheral nerves is visible throughout the tadpole. Credit: Taiyo Yamamoto, University of Zurich
Researchers at the University of Zurich have developed "Pythia", an AI system Named after the ancient Oracle at Delphi, that predicts how cells will repair DNA after CRISPR cuts, enabling dramatically more precise gene editing through AI-designed "molecular glue" templates. The tool has been successfully tested across human cells, frogs, and mouse brains, working even in non-dividing tissues like brain cells where traditional approaches struggle.
🧬💊 AI-designed antibiotics target superbugs
MIT researchers have used generative AI to design two new antibiotics atom-by-atom that successfully killed drug-resistant gonorrhoea and MRSA in laboratory and animal tests, moving beyond previous approaches that merely screened existing compounds. The AI analysed 36 million potential molecular structures to create entirely novel drugs, though they still require 1-2 years of refinement before entering the lengthy clinical trial process.
Narrative: starting to see Generative AI being applied to science… and the same acceleration we’ve observed in language models now emerging in DNA and biology. Darion Amodei’s compressed 21st century happening before our eyes?
🔮[Weak] signals
New tech that’s not AI…
Consumer tech
⭐📰Google launches "Preferred Sources" for personalised news results
While this gives users more control over their information diet, it also risks creating ideological bubbles that shield people from diverse perspectives on important topics.
Narrative: I’m tracking this closely as it’s pretty much what I’ve been building with Memia Sensorium: creating a curated list of (variably) trusted information sources and filters. Of course, Google's ultimate motivations are skewed by their ad- and attention- driven business model — so until they open-source their news feed algorithm they are no neutral information broker.
💾SIM-sized SSDs
Chinese manufacturer Biwin has developed "Mini SSD" storage cards that are roughly SIM-card sized (15mm x 17mm) but deliver blazing 3,700MB/s read speeds via PCIe 4x2 - nearly four times faster than the new MicroSD Express cards destined for Nintendo Switch 2. Two cutting-edge Strix Halo gaming portables from GPD and OneNetbook are already integrating dedicated slots for these tiny speed demons, which offer up to 2TB capacity with IP68 water resistance.
👓👄 Lip-reading AI smart glasses
UK researchers have developed AI-powered smart glasses that read lips to enhance hearing aids, using computer vision to filter speech from background noise and focus on individual speakers. The technology requires cloud processing but promises more affordable hearing assistance when it launches through manufacturers in 2026.
🚫German court revives copyright case against ad blockers
Germany's highest court has partially overturned a ruling in the 11-year Axel Springer vs Eyeo case, suggesting that ad blocking software could violate copyright law by modifying website code. The decision sends the case back to appeals court, potentially setting a precedent that could ban ad blockers in Germany—making it only the second jurisdiction after China to do so.
Fedora has officially added RISC-V as its fifth supported architecture, complete with dedicated Koji build infrastructure and ready-to-boot Fedora 41 images for popular boards like the StarFive VisionFive 2. This marks a significant milestone for the open-source instruction set architecture, which has been gaining traction since its Berkeley origins in 2010 but has suffered from limited mainstream OS support beyond Debian derivatives.
With a surge of new RISC-V hardware hitting the market, Fedora's commitment provides developers and enthusiasts with a regularly updated platform that actually keeps pace with the broader Linux ecosystem and accelerates the maturation of open hardware.
Robotics
Unitree’s H1 Humanoid Robot won gold at the first Humanoid Robot Games - completing the 1500m race in 7:10:
Narrative: I fully expect this to be faster than a human within 1 year.
The AI-powered robot can follow migrating herds at distances up to 2km without human interference, capturing behavioural data that would be impossible for researchers to gather due to the animals' natural wariness of humans.
Insta360's new drone startup Antigravity has unveiled the A1, a 249g folding drone that captures 8K 360-degree footage while letting pilots experience flight through immersive Vision goggles with head tracking. The system promises to eliminate the traditional disconnect between pilot and aircraft, using dual fish-eye lenses and "FreeMotion" technology to create a seamless first-person flying experience where you can look around naturally while the drone soars.
US FAA proposes new rule enabling nationwide drone deliveries
The FAA has proposed new rules allowing drones up to 600kg to fly beyond visual line of sight without individual waivers, potentially unlocking commercial drone delivery at scale across America. Major retailers like Walmart and Amazon are already positioning to expand their services, with Walmart planning 100-store coverage by mid-2026 and Alphabet-owned Wing promising sub-19-minute deliveries within 19km round trips.
The economics still look challenging though – at US$13.50 per package versus US$2 for traditional delivery … but that will get cracked with scale, particularly for hard-to-reach delivery locations.
🚁📦KARGO II
Pennsylvania's Piasecki Aircraft unveiled KARGO II, a VTOL drone capable of hauling 680kg of cargo autonomously to contested or remote locations. The system ships in a standard container and can be flight-ready within minutes, operated by non-aviation personnel with pinpoint delivery accuracy via multiple deployment methods.
🚁🛸⚡ Sparta: The drone 'mothership'
German manufacturer Quantum Systems is finalising development of "Sparta," a drone mothership being built in Ukraine that can carry and deploy multiple smaller drones at designated locations, effectively extending their operational range and reducing response times. The 8kg platform claims a 200km range, 6-8 hour endurance, and modular payload system, with serial production slated for late 2025.
✈️ China's VTOL jet drone: every warship an aircraft carrier
Photo via SCMP / Beihang University
SCMP / Beihang University
Chinese aerospace engineers from Beihang University have developed what appears to be the world's first high-speed vertical take-off and landing drone powered by a jet engine, capable of launching from ordinary warships even in rough seas. Unlike the US Air Force's runway-dependent XQ-58A Valkyrie, this platform combines vertical lift with jet-powered cruise flight in a single airframe, potentially transforming every Chinese destroyer or frigate into a mini aircraft carrier.
Ukrainian weapons manufacturer Fold is developing fibre-optic drones with a 100-kilometre range, doubling current capabilities to reach high-value targets that have moved farther from front lines as the battlefield evolves.
These drones use physical cables instead of radio frequencies, making them immune to electronic warfare jamming—with shotguns being soldiers' only real defence option.
The extended range comes with engineering trade-offs: heavier cables, reduced payload capacity, and increased snagging risks across the battlefield's "gray zone."
🚁👶🛡️Lithuania to train 22,000 people in drone skills
Lithuania is launching a €3.3 million programme to teach 22,000 people —including children as young as eight— how to build and operate drones as part of "civil resistance training" against potential Russian aggression. The curriculum scales from simple drone construction games for primary schoolers to advanced FPV drone manufacturing for teenagers, with training centres opening across the Baltic nation by 2028.
⚓Royal Navy controls robotic submarine from Australia
The Royal Navy's XV-Excalibur robotic submarine successfully patrolled Plymouth waters while being controlled from Australia 16,000 km away, demonstrating the future of autonomous naval warfare. This 12-metre XLUUV (Extra-Large Uncrewed Underwater Vessel) packs sophisticated sensors and equipment into a space traditionally reserved for human crew, representing Europe's most advanced uncrewed submarine.
Narrative: AUKUS nations demonstrating seamless trans-Pacific control of autonomous naval assets implies an emerging distributed maritime command structure — which could redefine naval presence and power projection in contested waters.
Biotech
🤖🌸Flowers designed for robots
Scientists from the Institute of Genetics and Developmental Biology (IGDB) of the Chinese Academy of Sciences have developed GEAIR (Genome Editing with Artificial-Intelligence-based Robots), a system that uses CRISPR gene editing to redesign flower structures so AI-controlled robots can perform cross-pollination 24/7, dramatically reducing breeding costs and timelines.
Narrative: Initial reaction is ew….. but is this agriculture's next evolutionary leap? From the Green Revolution's machinery-optimised crops to today's AI-optimised biology, literally redesigning nature's interfaces to unlock improvements in food system resilience.
🎯Nanobody 'drone strike' targets lung cancer precisely
Korean researchers have developed the A5 nanobody—a miniature antibody 10 times smaller than conventional ones—that precisely targets CD155 proteins on lung cancer cells and delivers chemotherapy drugs with surgical precision. Combined with liposomal drug delivery, the system achieved 70-90% tumour reduction in tests while leaving healthy organs completely unharmed, delivering three times more anticancer drugs to target cells than traditional methods.
🪰Scientists achieve first 'behaviour transplant' between species
Japanese researchers have achieved the world's first "behaviour transplant" between species, genetically rewiring fruit flies to perform completely foreign courtship rituals by manipulating a single gene.
By activating the Fruitless gene in specific neurons, they transformed singing D. melanogaster flies into “gift-giving” romantics like their evolutionary cousins D. subobscura—a behaviour that vanished from their lineage 35 million years ago.
The breakthrough suggests that dormant behaviours may lurk in our neural architecture, waiting for the right molecular switch to awaken ancient programs.
Slightly terrifying implications…🫣
🍄♻️Mycocycle uses mushrooms to eat trash and create bio-plastics
US startup Mycocycle is scaling its fungi-based waste processing system that literally trains white-rot mushrooms to consume construction waste, old tyres, and plastics (story from 1 year ago), transforming them into bio-based materials for insulation and concrete filler. The startup is targeting the US$624 billion plastics industry with a circular approach—”training” nature's recyclers to eat the very waste that traditional petrochemical companies produce.
Agriculture
🌱🥗Astronauts grow fresh food with new seed pillows
Crew-11 astronauts are now growing wasabi mustard greens, red Russian kale, and dragoon lettuce aboard the ISS using "seed pillows"—fabric pouches filled with baseball field clay and controlled-release fertiliser that work in microgravity. The VEG-03 experiments let astronauts choose their crops from a seed library, providing both nutrition and psychological benefits while testing agricultural systems for future Mars missions.
Stanford researchers have achieved a breakthrough by developing the first brain implant that can decode and vocalise "inner speech" - the words people think but don't attempt to say physically, reaching 74% accuracy in real-time translation for severely paralysed patients.
The system raises complex privacy questions since it can pick up unintended thoughts, leading the team to implement password protection using the phrase "chitty chitty bang bang" to prevent involuntary mind-reading.
This arrives in the same week as BCI investment heats up with OpenAI backing new venture Merge to compete with Neuralink.
Narrative: the dawn of practical telepathy? (complete with the need for mental firewalls?!)
Water infrastructure
💧 Water Farm 1: Subsea desalination goes deep
California is prototyping ocean-to-tap infrastructure with OceanWell's Water Farm 1 project, deploying 60 modular desalination pods 400 metres down in Santa Monica Bay to produce 227 million litres of drinking water daily by 2030.
The system uses crushing ocean pressure to drive reverse osmosis filtration, cutting energy consumption by 40% while filtering out everything from salt to PFAS "forever chemicals."
While the technology shows promise for addressing California's water crisis amid climate change and strained resources like the Colorado River, the project still requires feasibility studies and complex permitting processes before deployment.
Transport
Polestar 3 breaks EV range record with 935km journey
Polestar's flagship SUV has claimed the Guinness World Record for longest electric SUV journey on a single charge, covering 935 km (581.3 miles) across UK public roads in an unmodified production vehicle. The 23-hour trek exceeded the vehicle's official WLTP range of 438 miles by over 140 miles, achieving an impressive 5.13 miles per kWh efficiency in a 2.4-tonne SUV.
This milestone arrived just days after GM's Silverado EV unofficially hit 1,059 miles … range anxiety a thing of the past soon…
Energy
🏝️Indonesia launches massive 100GW village-scale solar program
Indonesia has launched an ambitious programme to deploy 100GW of solar PV and 320GWh of battery storage, with 80GW distributed across 80,000 villages through a new village cooperative model. The initiative aims to provide reliable electricity to rural areas while promoting economic development, though experts warn of significant challenges around skilled workforce availability and project coordination.
Crypto
Ethereum's future challenges
Cointelegraph's new documentary explores Ethereum's identity crisis as the 10-year-old blockchain grapples with layer-2 fragmentation that's improved user experience but gutted validator revenues by 90%. The film features key ecosystem players debating whether Ethereum should compete directly with faster chains like Solana or double down on its strengths as a decentralised settlement layer for real-world assets and institutional adoption.
Despite operational momentum and SEC lawsuit dismissal providing regulatory clarity, HNT token price fell 25.6% as the network underwent its third halving, cutting annual emissions from 15 million to 7.5 million tokens.
Narrative: the disconnect between surging utility metrics and declining token value suggests the market hasn't fully grasped how DePIN networks create value through actual infrastructure usage rather than speculative trading.
(Disclosure: I’m a long-suffering HNT HODLer…)
⏳ Zeitgeist
Once around the world outside tech… keeping it brief and treading lightly…
Human wellbeing
New study challenges World Bank's extreme poverty decline claims
Using a "Basic Needs Poverty Line" that accounts for actual survival costs, researchers found poverty only dropped from 23% to 17% between 1980-2011, while the absolute number of extremely poor people actually increased from 1.01 billion to 1.20 billion.
The findings suggest that market reforms of the 1980s-90s may have pushed an additional billion people into extreme poverty.
Pollution
♻️EU pushes for plastic pollution treaty amid deadlock
As covered over the last couple of issues, the EU is still pushing for a global plastic pollution treaty in Geneva talks ending Thursday, but faces resistance from oil-producing nations who prefer focusing on waste management rather than production limits. With plastic production set to triple by 2060 and five previous negotiation rounds having failed over 2.5 years, Danish Environment Minister Magnus Heunicke warns there's "going to be a whole lot more drama" ahead. The deep tension between petrochemical economics and environmental necessity is playing out in real-time.
Climate change
Tracking more extreme weather phenomena…
🔥🌡️🇪🇺Europe faces record heat as wildfires rage
As I can attest, record-breaking temperatures (above 40°C in many places) are triggering widespread wildfires across Europe, from Spain to Turkey, forcing thousands to evacuate as the continent warms twice as fast as the global average since the 1980s.
Emergency services are stretched thin battling blazes that have killed at least one person and hospitalised dozens more from smoke exposure.
🌀Hurricane Erin threatens US coast
Hurricane Erin underwent one of the fastest intensification cycles on record, briefly hitting Category 5 before settling as a massive Category 3 storm now menacing the Caribbean and US East Coast with 6-metre waves and flooding despite staying offshore. Climate scientists point to ocean temperatures made up to 100 times more likely by climate change as the fuel for this "heat engine" that's rewriting the playbook on hurricane behaviour. The storm's unusually large footprint—hurricane-force winds extending 130km from centre—offers a preview of how our warming oceans are manufacturing bigger, nastier storms
🏔️💧Alaska's capital evacuates as glacier unleashes record flooding
Alaska's capital Juneau1 is evacuating 460 homes as the Mendenhall Glacier releases a record-breaking 15 billion gallons of meltwater—equivalent to 23,000 Olympic swimming pools—into the Mendenhall River. The city has deployed 10,000 sand-filled Hesco barriers along 2.5km of riverbank as a temporary levee, but these glacial outburst floods have become an annual summer event since 2011 and are expected to continue for another 25-60 years.
Alaska is warming twice as fast as the rest of the US: this is climate change in real-time.
🌳⚠️🔥 Amazon rainforest tipping points
New research identifies three critical thresholds that could flip the Amazon from rainforest to savanna within a century: 65% forest cover loss, 10% reduction in Atlantic moisture, or 6% rainfall decrease. The study reveals how the forest's water recycling system—where trees create the rain they need to survive—becomes a death spiral once disrupted, with deforestation and climate change working as deadly multipliers.
Health
🦠Mpox spreads rapidly across Congo
Mpox has evolved from a rural wildlife spillover disease to a human-adapted pathogen spreading across 24 of 26 provinces in the Democratic Republic of Congo, with a new subclade Ib variant showing sustained human-to-human transmission among adults. The virus has now reached Kinshasa, a 17-million-person transport hub connecting much of Africa, while also demonstrating mother-to-baby transmission during pregnancy. As one researcher bluntly put it: "Viruses don't have passports" – a reminder that our interconnected world makes local outbreaks everyone's problem.
Vegetarian diets slash cancer risk by 45%
A comprehensive 13-year study tracking nearly 80,000 Seventh-day Adventists (huh?) found vegetarian diets reduced stomach cancer risk by 45% and lymphomas by 25%, with an overall 12% cancer risk reduction. The research suggests the protective effects are strongest for gastrointestinal cancers, where food directly contacts digestive organs, though with the skewed sample set we're still looking at correlation rather than definitive causation.
Infrastructure
🏗️💸🇸🇦Saudi's PIF writes down $8 billion on megaprojects
Reality check: Saudi Arabia's Public Investment Fund has written down US$8 billion from its five mega-projects as costs spiral and oil prices dip, with flagship NEOM project now under review for potential job cuts and scope reductions.
Despite the setback, PIF still expects to more than double its assets to US$2.67 trillion by 2030, pivoting heavily toward domestic investments (now 80% of portfolio) while creating 16 new companies including AI mega-fund HUMAIN and aerospace firm Neo Space.
Nigeria's decade-long quest for a 4,000km high-speed rail network connecting Lagos to Port Harcourt has cleared a major hurdle, with local firm De-Sadel Nigeria and Chinese financier China Liancai Petroleum Investment Holdings presenting formal proof-of-funds for the US$60 billion project. The ambitious infrastructure play promises to open completed sections within three years, while also converting existing diesel trains to run on liquefied and compressed natural gas.
Russian hackers hijacked Norwegian dam, spilled water
Russian hackers allegedly hijacked a Norwegian dam in April, opening floodgates for four hours and releasing the equivalent of three Olympic swimming pools worth of water before being stopped. Norway's spy chief publicly blamed Russian hackers for the Bremanger dam breach, though Russia predictably denied involvement.
Miscellaneous
🗺️🌍📏 Correct the Map: Africa's cartographic rebellion
Mercator Projection (left) and “Equal Earth” Projection (right)
The African Union has officially backed the "Correct the Map" campaign, which challenges the centuries-old Mercator projection that makes Africa appear significantly smaller than its actual 30.3 million square kilometres.
AU Deputy Chairperson Selma Malika Haddadi argues this isn't just about cartography—it's about correcting a visual narrative that renders Africa "marginal" on the global stage.
🎭💼Pretend To Work
Young unemployed Chinese adults are paying US$4-7 daily to sit in fake offices and pretend to work, as youth unemployment exceeds 14% and real jobs become scarce. These mock workplaces provide computers, meeting rooms, and the social structure of employment while users job hunt, develop skills, or maintain dignity in front of family and universities.
Narrative: With AI, the boundaries between “work” and “pretend to work” could melt away very quickly… are these “co-pretending” spaces actually part of our future social infrastructure?
🧠Mind expanding
A Phylogeny of Agents
Fascinating paper from Jonas Hallgren, markov from Equilibrium, proposing a research agenda for exploring A Phylogeny of Agents:
“…[For a given AI firm, there are] different ways that alignment and agency could work from different intentional stances:
From an economic perspective, the AI firm might be perfectly compliant. It follows market regulations, responds to price signals, allocates resources efficiently, and maximizes shareholder value. The firm appears well-aligned with economic incentives and legal frameworks. Economists studying this system would conclude it's behaving as a rational economic agent should.
From a decision theory perspective, the firm might be executing its stated utility function flawlessly. Its sub-agents optimize for clearly defined objectives, exhibit goal-directed behavior toward specified targets, and demonstrate adaptive learning within their designed parameters. AI researchers examining the system would find textbook alignment between the AI's behavior and its programmed goals.
From a cooperative AI perspective, this same firm might be generating catastrophic coordination failures. Despite following its individual incentives perfectly, it could be contributing to race dynamics, undermining collective welfare, or creating systemic risks that no individual actor has incentive to address. Researchers studying multi-agent dynamics would see dangerous misalignment at the system level.
From a biological systems perspective, the firm might be optimized for short-term efficiency but catastrophically fragile to environmental shocks. Like a monoculture lacking genetic diversity, it could be heading toward collapse because it lacks the redundancy, adaptability, and resilience mechanisms that biological systems evolved for survival under uncertainty.
This is where the phylogeny of agents research program becomes useful. Darwin's phylogenetic tree revolutionized biology by revealing the evolutionary relationships between species. If we map the phylogeny of analytical approaches to agency, we could transform how we understand and align complex intelligent systems.”
What a potential phylogeny of intentional stances might look like.
Paris Conversations with Yann Lecun
An engaging and reflective interview with Meta’s phlegmatic AI Chief Scientist, back in Paris where he started out:
🕳️Webb telescope discovers oldest black hole ever found
The James Webb Space Telescope has spotted the most distant confirmed black hole yet, lurking inside a peculiar early galaxy dubbed a "Little Red Dot" that existed just 500 million years after the Big Bang. This cosmic heavyweight—up to 300 million times our sun's mass—is so massive relative to its galaxy that it's forcing astronomers to rethink how quickly black holes could bulk up in the universe's infancy. The discovery suggests either black holes grew impossibly fast in the early cosmos, or they started out far chunkier than our models predict—both scenarios that would rewrite the textbook on cosmic evolution.
📎Paperclip-sized spacecraft could explore nearby black holes
The audacious mission would take roughly 70 years to reach its target and aims to probe the extreme physics where Einstein's equations break down.
The plan faces minor hurdles like needing technology that doesn't yet exist, finding a black hole closer than the current nearest one at 1,500 light years away, and securing over a trillion dollars in funding.
💭Meme stream
A lot of eclectica filling up my feed this week!
☄️70 years ago - meteorite strikes human
TIL: Alabama woman Ann Hodges remains the only documented person ever struck by a meteorite, surviving a 4.5-billion-year-old space rock that crashed through her roof in 1954 — over 70 years ago— leaving her with just a bruised hip (and unwanted fame). A recent near-miss in Georgia, where meteorite fragments missed a resident by just 14 feet, shows how extraordinarily rare these cosmic encounters remain—with odds compared to being simultaneously hit by a tornado, lightning, and hurricane.
🏴☠️Phrack magazine celebrates 40 years of hacker culture
Phrack Magazine released its 72nd issue, marking an extraordinary 40-year run as the underground hacker scene's most enduring publication. The legendary zine continues to document and shape hacker culture, bridging generations of security researchers with technical papers, exploit techniques, and cultural commentary that spans from phone phreaking to modern cloud security. Lots of fascinating rabbit holes to go down...
🖥️☕🚫Starbucks Korea bans desktop PCs
While I’m enjoying laptop-friendly European cafes… Starbucks South Korea has banned desktop computers, printers, and privacy partitions after customers began treating coffee shops as literal home offices, complete with full PC setups:
Photo via TechSpot
The "cagongjok" (café study tribe) phenomenon has pushed the boundaries of what constitutes reasonable café workspace etiquette.
Narrative: Apparently US$3 coffee covers just 1 hour 42 minutes of “profitable seat time” before customers become a net loss. Someone definitely needs to create a better business model for seamless nomadic working than coffee and croissants…
The future is weird, y’all …. my AI bubble on X can’t quite decide if this meticulously dated Oasis AI Karaoke is AI-generated or not…
Or this… Nirvana covered like you’ve never seen it before:
Gold either way, I reckon… no way to tell pretty soon. If it’s AI, it’s high art.
Spitting Image to the rescue
I grew up with the Spitting Image puppets in the UK… a safety pressure valve on politician and “celebrity” egos back in the 90s until it faded out with the last episode in 1996. The series returned briefly in the early 2020s…but then was discontinued again. And now it’s returned to YouTube as “The Rest Is Bullshit” … and judging from the first episode, this year’s series takes no prisoners…
(Loving Paddington Bear with a Peruvian accent — obviously, when you think about it… — and with a taste for “South American Product”…LOL.)
Warning: completely NSFW … goes way beyond the boundaries of what most people would consider good taste … but a welcome no-holds-barred lampooning of Trump and many others who deserve it. More like this please.
Vincent’s tale
Finally, more modern poetry / virtuoso guitar playing / unhinged stream-of-consciousness narcissism (you decide) from the hardy perennial Brighton busker Ren. Super talented musician… more original than most of the content you’ll see on YouTube these days…
🙏🙏🙏 As always thank you to everyone who takes the time to get in touch with links and feedback. Appreciated!
