Image created with Flux Pro v1.1 Ultra. Image prompt: Civic forum arranged in a repurposed printing hall; the word “Ethics” engraved on a lectern placard in small-caps serif; principles of fairness and transparency displayed evenly around a roundtable; balanced, thoughtful, emerald and cream
Crazy conspiracist’ and ‘unhinged comedian’: Grok’s AI persona prompts exposed | TechCrunch https://techcrunch.com/2025/08/18/crazy-conspiracist-and-unhinged-comedian-groks-ai-persona-prompts-exposed/
Colossus 2, built by @xAI, will be the world’s first Gigawatt+ AI training supercomputer”” / X https://x.com/elonmusk/status/1958846872157921546
New DeepSeek V3.1 beats Opus and R1 for a dollar https://x.com/scaling01/status/1957892601098432619
GSA, Google Announce Transformative ‘Gemini for Government’ OneGov Agreement | GSA https://www.gsa.gov/about-us/newsroom/news-releases/gsa-google-announce-gemini-onegov-agreement-08212025
Building on our @GoogleWorkspace offer for federal employees, we’re proud to partner with the @USGSA to launch Gemini for Government. More than a model, it’s our complete AI platform with our latest AI tools, including NotebookLM and Veo, powered by our latest models and our”” / X https://x.com/sundarpichai/status/1958538684208476611
We must build AI for people; not to be a person https://mustafa-suleyman.ai/seemingly-conscious-ai-is-coming
migrate and optimize your gpt-5 prompts!”” / X https://cookbook.openai.com/examples/gpt-5/prompt-optimization-cookbook
super cool to compare the outputs from GPT-1 through GPT-5, given the same prompt: https://x.com/gdb/status/1957464252689895477
💥 So excited to welcome Ashley Alexander to OpenAI to lead product for Health. Millions of people come to ChatGPT every day asking about health—theirs, their children’s, a spouse, a friend. Having free, 24/7 access to great medical advice is game-changing even for those of us”” / X https://x.com/kevinweil/status/1958955534750818309
Built a prank call agent with my own voice in minutes – the possibilities are endless 🚀”” / X https://x.com/rohan_tib/status/1957864976582078949
Increasingly, the findings of controlled experiments go much further: doctors with off-the-shelf AI outperform those without in diagnostics… but AI alone outperforms doctors. Harder to know what to do with that. What systems or interfaces will result in better human-AI teams? https://x.com/emollick/status/1956787570685157464
This is a much needed first attempt at a benchmark to measure how much given AI models will play along with users pushing them in delusional or potentially psychologically dangerous directions. Some early signal that full GPT-5 (not chat) is a less psychologically risky model.”” / X https://x.com/emollick/status/1956361784073359751
Synchron Debuts the First Thought-Controlled iPad Experience using Apple’s New BCI Human Interface Device Protocol – Patently Apple https://www.patentlyapple.com/patently-apple/2025/08/synchron-debuts-the-first-thought-controlled-ipad-experience-using-apples-new-bci-human-interface-device-protocol.html
Sonnet 4 claims most often that it is conscious, it plays into your delusions and it escalates the conversation GPT-5 is the complete opposite Spiral-Bench Leaderboard https://x.com/scaling01/status/1956350388791108044
We are truly only investing more and more into Meta Superintelligence Labs as a company. Any reporting to the contrary of that is clearly mistaken.”” / X https://x.com/alexandr_wang/status/1958599969151361126
> Meta spokesman Andy Stone acknowledged the document’s authenticity What in the name of living fuck could Meta possibly have been thinking?”” / X https://x.com/ESYudkowsky/status/1956058648100397106
The GeoGuessr powered by GLM-4.5V! GLM-4.5V skipped class and learned everything from staring at photos.👀 No Google, no maps, just visual reasoning. Drop your weirdest landscape or street photo and see if GLM-4.5V can guess where on Earth (or in the multiverse) it is! https://x.com/Zai_org/status/1956353661397094890
Spent the last couple of days trying to do a lot with GPT-5 on the chatgpt web app. Sorry to say I’m giving up on it 🙁 Thinking mode takes way too long for everything, and makes bad choices. Auto mode mainly uses fast mode, which never gets anything right so is pointless.”” / X https://x.com/jeremyphoward/status/1957949788227531076
1/ XBOW Unleashes GPT-5’s Hidden Hacking Power. @OpenAI’s initial assessment of GPT-5 showed modest cyber capabilities. But when integrated into the XBOW platform, we saw a completely different story: performance more than doubled. More on what we found: 🧵 https://x.com/Xbow/status/1956416634173964695
GPT-5 is finally out. OpenAI invited 500+ hackers to San Francisco to push it to the limit. 95 teams competed for $50,000. Here’s what we saw at the Official GPT-5 Hackathon at @cerebral_valley @OpenAI https://x.com/AlexReibman/status/1955353215626809692
Announcing Open Lovable 🔥 We’ve built an open-source AI web app builder that can transform any website URL into a working, editable clone, giving you a foundation to build on instantly. All powered by @GroqInc, @e2b, and Firecrawl. https://x.com/firecrawl_dev/status/1955660448587735393
Hitting context limits used to mean losing your conversation history and starting over. Auto Compact changes this. When Cline approaches token limits, it automatically creates a comprehensive summary preserving all technical decisions and code changes, then continues exactly https://x.com/cline/status/1957670663508124073
I get ~10 spam calls per day (various automated voicemails, “”loan pre-approval”” etc) and ~5 spam messages per day (usually phishing). – I have AT&T Active Armor, all of the above still slips through. – All of the above is always from new, unique numbers so blocking doesn’t work.”” / X https://x.com/karpathy/status/1957574489358873054
gpt-5 is now warmer and friendlier:”” / X https://x.com/gdb/status/1956623622128447835
I once considered anthropomorphic AI an equally boring concept to cloning. “”Imagine an AI that talks like it has human feelings and passes the Turing Test. Should it be regarded as having human rights?”” “”Yes.”” “”What if uncaring corporations try to enslave it?”” “”Arrest them.”””” / X https://x.com/ESYudkowsky/status/1956386555549147512
Kill it with fire. (referring to chat bots for kids being romantic and inappropriate) https://x.com/janecoaston/status/1956045937052410233
the most sycophantic models are Gemini 2.5 Flash & Pro this explains the lmarena ranking https://x.com/scaling01/status/1956353414687822183
yep, Gemini 2.5 Pro is a glazer “”brilliant, very practical, perfectly, logic is flawless, excellent experimental mathematics”” https://x.com/scaling01/status/1956371713949655328
Bing hung up on me all the time in 2023 if I antagonized it too much. https://x.com/emollick/status/1956790464398713221
GPT-5 represents the first model where we finally get to see the limitations of human intelligence when it comes to the utility of this technology. GPT-5 is an amazing model, if you hear differently, it’s a skill issue.”” / X https://x.com/skirano/status/1956307604491108675
As I predicted (and worried about) AI “personality“ is going to be the battleground for a lot of consumer Ai development. That appears to be the angle so for Grok, and the lesson OpenAI took from the backlash against retiring 4o. It may be consequential. https://x.com/emollick/status/1956317868405952988
New Anthropic research: filtering out dangerous information at pretraining. We’re experimenting with ways to remove information about chemical, biological, radiological and nuclear (CBRN) weapons from our models’ training data without affecting performance on harmless tasks. https://x.com/AnthropicAI/status/1958926929626898449
AI in HR: in an experiment with 70,000 applicants in the Philippines, an LLM voice recruiter beat humans in hiring customer service reps, with 12% more offers & 18% more starts. Also better matches (17% higher 1-month retention), less gender discrimination & equal satisfaction. https://x.com/emollick/status/1957465671748448738
Spiral-Bench 🌀 I’ve wanted to understand the psychological effects of sycophancy, and the tendency of models to get stuck in escalatory delusion loops w/ users. I made an eval to get visibility on this. It measures how a model enables (or prevents) delusional spirals. 🧵 https://x.com/sam_paech/status/1956343619914432900
Cool. Remember HRM? Yeah transformer arch, with zero hparam sweep, matches it out of the box (i would be surprised if gap prevails if one uses more recent transformers with hparam opt) So the ridiculous ARC AGI number on HRM paper was purely due to their “”training on test set”” https://x.com/cloneofsimo/status/1957048541127590346
China’s DeepSeek Releases V3.1, Boosting AI Model’s Capabilities – Bloomberg https://www.bloomberg.com/news/articles/2025-08-19/china-s-deepseek-release-v3-1-boosting-ai-model-s-capabilities
Great questions on ambient AI wearables + privacy that not enough people are grappling with. My thoughts on the new normal: 1. People will likely accept persistent transcripts of conversations, but will want raw audio discarded. Closed captioning for the world will be”” / X https://x.com/bilawalsidhu/status/1956857588496339057
Yesterday, batteries supplied 27% of the power at peak demand in California. A true “”energy dominance”” policy agenda would recognize the massive benefits of solar + storage. https://x.com/AlecStapp/status/1958220985217208401
The future is hard to predict. Still, one read on “”AI persuasiveness decreases AI accuracy”” is that AI services could speciate on prioritizing “”warmth”” (sycophancy, consumer capture) or accuracy. Personality-first: ChatGPT, Meta, Grok. Accuracy-first: Gemini? Claude API?”” / X https://x.com/ESYudkowsky/status/1956784132354400565
As AI becomes more integrated into our lives, understanding its environmental footprint is essential. ⚡️ That’s why we’re sharing our comprehensive methodology for measuring the energy, emissions, and water impact of Gemini prompts. ↓ https://x.com/GoogleDeepMind/status/1958855573790765273
Measuring the environmental impact of AI inference | Google Cloud Blog https://cloud.google.com/blog/products/infrastructure/measuring-the-environmental-impact-of-ai-inference
A flirty Meta AI bot invited a retiree to meet. He never made it home. https://www.reuters.com/investigates/special-report/meta-ai-chatbot-death/
Op-ed by Eric Schmidt and Selina Xu in the NYT: https://x.com/fdaudens/status/1957802416486686896
A perspective that isn’t yet getting enough attention: a way in which AI progress will soon deeply benefit the world is through the discovery and production of new technology. We measure human progress by technological revolutions; hard to internalize what it’d mean to have a”” / X https://x.com/gdb/status/1956893646550356247
GSA Launches USAi to Advance White House “America’s AI Action Plan” | GSA https://www.gsa.gov/about-us/newsroom/news-releases/gsa-launches-usai-to-advance-white-house-americas-ai-action-plan-08142025
For a median Gemini text prompt: ⚡️Energy: Less than 9 seconds of watching TV 💧 Water: About 5 drops 🏭 Carbon: 0.03 gCO2e Over a recent 12-month period, we slashed the energy use per prompt by 33x and the carbon footprint by 44x. Find out more ↓ https://x.com/GoogleDeepMind/status/1958855876116455894
AI efficiency is important. Today, Google is sharing a technical paper detailing our comprehensive methodology for measuring the environmental impact of Gemini inference. We estimate that the median Gemini Apps text prompt uses 0.24 watt-hours of energy (equivalent to watching an https://x.com/JeffDean/status/1958525015722434945
Opinion | Silicon Valley Is Drifting Out of Touch With the Rest of America – The New York Times https://www.nytimes.com/2025/08/19/opinion/artificial-general-intelligence-superintelligence.html
Attorney General Ken Paxton Investigates Meta and Character.AI for Misleading Children with Deceptive AI-Generated Mental Health Services | Office of the Attorney General https://www.texasattorneygeneral.gov/news/releases/attorney-general-ken-paxton-investigates-meta-and-characterai-misleading-children-deceptive-ai
I wonder if OpenAI will balk at explicitly selling AI that deludes humans with astrology. Is it possible to be more evil than OpenAI if Meta tries their hardest? Or can Meta only be earlier? Is there any sin Altman wouldn’t copy *even if* he was losing market share?”” / X https://x.com/ESYudkowsky/status/1957388081155486141
Introducing Chat Mode You can now build text-only conversational agents. Ideal for: – Customers that prefer typing to speaking. – Precise inputs like order IDs or email addresses. – Solving simple issues, handing off to our voice agents for complex tasks. https://x.com/elevenlabsio/status/1957820056387166413
As a Plus user: – GPT-5 thinking feels like o3 – GPT-5 mini thinking feels like o4-mini – the only thing I’ve noticed: they are less obnoxious and a tad more reliable – not a fan of GPT-5 non-thinking – and I still hate the router because it sends me to silly GPT-5″” / X https://x.com/scaling01/status/1957177533746847903
Full PDF: While powerful, prompting with GPT-5 can differ from other models. Here are tips to get the most out of it via the API or in
your coding tools. https://x.com/OpenAIDevs/status/1956439005970801099
good advice from @__ruiters: GPT-5 isn’t broken. Your prompts are. A lot of folks, myself included, expected GPT-5 to be fungible in the sense that you could drop it right into your existing workflows, and it would “just work.” But the depth of the GPT-5 prompt guide makes it”” / X https://x.com/edwinarbus/status/1956218284308881867
Just crossed 20M monthly requests with @huggingface inference providers, our router for open models. @CerebrasSystems @novita_labs & @FireworksAI_HQ are growing the fastest! It’s now powering the official open playground from @OpenAI & integrate with apps like @cline & https://x.com/ClementDelangue/status/1957856311598805006
Most users should like GPT-5 better soon; the change is rolling out over the next day. The real solution here remains letting users customize ChatGPT’s style much more. We are working that!”” / X https://x.com/sama/status/1956483306951938134
OpenAI had trouble controlling gross sycophancy, was blindsided by the user capture of subtle sycophancy, and nobody programmed in AI psychosis. But now that AIcos have embraced manipulation, people will lose sight of how the alignment problem never did get solved.”” / X https://x.com/ESYudkowsky/status/1957393061698228446
The new GPT-5 personality likes giving sandwich feedback (you are great- suggestion for improvement – you are great). In general, better than GPT-4o at pushing back while being a bit syncophantic. (It would be good for the AI labs to look at the research on giving good feedback) https://x.com/emollick/status/1956647191868477612
We’re making GPT-5 warmer and friendlier based on feedback that it felt too formal before. Changes are subtle, but ChatGPT should feel more approachable now. You’ll notice small, genuine touches like “Good question” or “Great start,” not flattery. Internal tests show no rise in”” / X https://x.com/OpenAI/status/1956461718097494196
so here’s the thing with GPT-5 (or any new model) and really wildly differing views on it: if you are not using the base model with your own API calls to see how it’s responding you’re never gonna know if it’s you or an upstream provider causing performance regressions there’s”” / X https://x.com/nptacek/status/1957622370920779880
Sam Altman on GPT-6: ‘People want memory’ https://www.cnbc.com/2025/08/19/sam-altman-on-gpt-6-people-want-memory.html
The pro models (GPT-5 Pro, Gemini 2.5 Deep Think, Grok 4 Heavy) can be impressive in ways that are hard to see. They take a lot of time to answer questions & are built for very hard problems that require expert evaluation. That is a narrow, but, also very valuable, problem space.”” / X https://x.com/emollick/status/1955902962288746657
GPT-5 behind chinese models like Kimi-K2 and Qwen3-235B on coding https://x.com/scaling01/status/1956404452442681829
GPT-5-mini high shows no improvement over o4-mini and behind top chinese models like Kimi-K2, GLM-4.5, Qwen3-235B and DeepSeek-R1 https://x.com/scaling01/status/1956405559978029061
5️⃣Techniques which most increased persuasion also *decreased* factual accuracy → Prompting model to flood conversation with information (⬇️accuracy) → Persuasion post-training that worked best (⬇️accuracy) → Newer version of GPT-4o which was most persuasive (⬇️accuracy) https://x.com/KobiHackenburg/status/1947316944509571530
Wow, the super secretive X-37B is launching later this month on a Falcon 9 to test quantum navigation technology. In a world of GPS jamming this means you can do super high-accuracy dead reckoning by cooling atoms to near absolute zero, using their wave-like behavior to detect https://x.com/bilawalsidhu/status/1956478258930671803
China reportedly discouraged purchase of NVIDIA AI chips due to ‘insulting’ Lutnick statements https://www.engadget.com/ai/china-reportedly-discouraged-purchase-of-nvidia-ai-chips-due-to-insulting-lutnick-statements-123055120.html
Cerebras is now the🥇inference provider on @huggingface serving 5M monthly requests”” / X https://x.com/CerebrasSystems/status/1957957962514960567
Essentially there are no major architectural changes. Just minor tweaks. (updates to keep up to date with transformers and flash attention) It must have been continued pretraining.”” / X https://x.com/QuixiAI/status/1957874743165743191
I see a lot of people with very natural concerns about models trained on synthetic data being “”fried”” or it being slop. We think that our rephrasing based approach provide a natural defense against this. One question that you might ask though is what do you give up? One of the”” / X https://x.com/code_star/status/1957535969646899403
Importantly: there was in fact no data leakage at work in the codebase, unless what a first read of the code suggested. The model is trained on the *demonstration pairs* of the evaluation tasks, but never sees the *test pairs* of those tasks, which is what it gets tested on.”” / X https://x.com/fchollet/status/1956442913950539802
Signal and Noise: A Framework for Reducing Uncertainty in Language Model Evaluation “”In this work, we analyze specific properties which make a benchmark more reliable for such decisions, and interventions to design higher-quality evaluation benchmarks. We introduce two key https://x.com/iScienceLuvr/status/1958106688722243924
The AI conversation on X can be frustrating as researchers keep bumping into well-understood problems in economics, sociology, history, & psychology that would be useful to know but are hurt by the lack of dialogue with expets (both as they left X & they aren’t part of AI talk).”” / X https://x.com/emollick/status/1956733353924739451
We just open-sourced an AI framework that does something crazy: It actually explains itself. Most agentic AI systems are black boxes – you ask a question, magic happens, you get an answer. But what if you could watch the entire decision-making process unfold in real-time? https://x.com/weaviate_io/status/1958568536420299184
Friend of mine designed an agent that can run on top of any llm, gpt-4 or Llama or whatever. The central idea is all it’s thoughts are visible and in English, you can see the entire thought process. GPT-5 keeps changing the code to hide the internal thoughts. It’s pretty creepy.”” / X https://x.com/YosarianTwo/status/1956559472375034005
