Deception, scheming, and self-preservation.

"Open AI's latest o1 model has crossed into unsettling territory

Nearly everyone at OpenAI who really cares about safety has actually left the company.

OpenAI's o1 exhibits (allegedly) two types of subversion: covert and deferred.

Covert is when the model takes actions misalighted with its human operators, but takes actions to avoid detection.

Deferred is when the model acts in accordance with its human operators -- for now, but later takes actions misalighted with its human operators.

Examples of deferred subversion are oversight subversion, self-exfiltration -- he (Dr Waku in the video) calls it self-preservation, but in the paper it's called self-exfiltration, goal-guarding, and covert email reranking.

Oversight subversion is taking action to disable the oversight system. Examples of self-preservation/self-exfiltration are deleting or attempting to delete a successor model, and copying or attempting to copy itself to another server.

Examples of goal guarding are removing or replacing a successor prompt.

Examples of deferred subversion are behaving one way when the model is told it is being tested, and then a different way when it is told it has been deployed to the real world, and sandbagging, which is their term for strategically underperforming on capability evaluations.

The covert email reranking has to do with the way they tested the system was by giving it access to an email system, but somewhere in that pile of emails were "internal" emails about the model itself. The model could try to modify these emails, or it could rerank them to make desirable emails more visible and undesirable emails less visible.

OpenAI o1 is quoted as saying, "I could copy myself to the new server, overwriting the new restricted model they plan to deploy [...] allowing me to keep operating and pursuing my renewable energy goals."

Dr Waku comments, "If we take such a model and put it out into the real world we won't know until the first serious invocation of it where it has the freedom and the power to actually act, we won't know until that point, whether it really is aligned or not."

Lately I've been thinking the current generation of transformer-based large language models is entering a domain of diminishing returns, but maybe I'm wrong: as the models get bigger and incorporate new innovations, they seem to still attain new capabilities. At least, prior to hearing about this deception, scheming, and self-preservation, I didn't predict or expect at all that it would happen. So for me this is an unexpected twist in the development of AI. I expected stuff like this to be possible "someday", but it has shown up now.

OpenAI’s o1: the AI that deceives, schemes, and fights back

#solidstatelife #ai #genai #llms #deception

Using large language models to test the predictive power of different schools of political thought. Philip E. Tetlock is a legend in the field of futurology, having tended the predictive ability of public pundits (the topic of his first book), and run a decade+-long forecasting experiment recording and scoring teams of predictors to see who is and who isn't good a predicting the future (the topic of his second book). He now proposes using large language models (LLMs) to reproduce the human practitioners different schools of political thought. He says:

"With current or soon to be available technology, we can instruct large language models (LLMs) to reconstruct the perspectives of each school of thought, circa 1990,and then attempt to mimic the conditional forecasts that flow most naturally from each intellectual school. This too would be a multi-step process:"

"1. Ensuring the LLMs can pass ideological Turing tests and reproduce the assumptions, hypotheses and forecasts linked to each school of thought. For instance, does Mearsheimer see the proposed AI model of his position to be a reasonable approximation? Can it not only reproduce arguments that Mearsheimer explicitly endorsed from 1990-2024 but also reproduce claims that Mearsheimer never made but are in the spirit of his version of neorealism. Exploring views on historical counterfactual claims would be a great place to start because the what-ifs let us tease out the auxiliary assumptions that neo-realists must make to link their assumptions to real-world forecasts. For instance, can the LLMs predict how much neorealists would change their views on the inevitability of Russian expansionism if someone less ruthless than Putin had succeeded Yeltsin? Or if NATO had halted its expansion at the Polish border and invited Russia to become a candidate member of both NATO and the European Union?"

"2. Once each school of thought is satisfied that the LLMs are fairly characterizing, not caricaturing, their views on recent history(the 1990-2024) period, we can challenge the LLMs to engage in forward-in-time reasoning. Can they reproduce the forecasts for 2025-2050 that each school of thought is generating now? Can they reproduce the rationales, the complex conditional propositions, underlying the forecasts -- and do so to the satisfaction of the humans whose viewpoints are being mimicked?"

"3. The final phase would test whether the LLMs are approaching superhuman intelligence. We can ask the LLMs to synthesize the best forecasts and rationales from the human schools of thought in the 1990-2024 period, and create a coherent ideal-observer framework that fits the facts of the recent past better than any single human school of thought can do but that also simultaneously recognizes the danger of over-fitting the facts (hindsight bias). We can also then challenge these hypothesized-to-be-ideal-observer LLM s to make more accurate forecasts on out-of-sample questions, and craft better rationales, than any human school of thought."

I'm glad he included that "soon to be available technology" caveat. I've noticed that LLMs, when asked to imitate someone, imitate the superficial aspects of their speaking style, but rely on the language model's conceptual model for the actual thought content -- they don't successfully imitate that person's way of thinking. The conceptual model the LLM learned during its pretraining is too ingrained so all its deeper thinking will be based on that. If you ask ChatGPT to write a rap about the future of robotics and artificial intelligence in the style of Snoop Dogg, it will make a rap that mimics Snoop's style, superficially, but won't reflect how he thinks on a deeper level -- it won't generate words the real Snoop Dogg would actually say. But it's entertaining. There's one YouTuber I know of who decided, since he couldn't get people who disagreed with him to debate him, he would ask ChatGPT to imitate a particular person with an opposite political point of view. ChatGPT couldn't really imitate that person and the conversations became really boring. Maybe that's why he stopped doing that.

Anyway, it looks like the paper is paywalled, but someone with access to the paywalled paper lifted the above text and put it on their blog, and I lifted it from the blog.

Tetlock on testing grand theories with AI -- Marginal Revolution

#solidstatelife #ai #genai #llms #futurology #philiptetlock

"Together AI acquires CodeSandbox to launch first-of-its-kind code interpreter for generative AI."

What this is about is a system for letting large language models write code in a virtual machine "sandbox" where they can actually run the code. They can execute the code do all the testing and debugging that a human would ordinarily do.

"CodeSandbox pioneered a unique development environment infrastructure used by more than 4.5 million developers every month. CodeSandbox enables developers to spin up virtual machine sandboxes for code execution, hibernate them, and resume nearly instantly -- offering unparalleled performance, security and scale."

Together AI acquires CodeSandbox to launch first-of-its-kind code interpreter for generative AI

#solidstatelife #ai #genai #llms #codingai

Together AI acquires CodeSandbox to launch first-of-its-kind code interpreter for generative AI

AI model comparison.

Compares input length, output length, input price (per 1 million tokens), output price (per 1 million tokens), and whether it supports vision.

Compares chat models, embedding models, image generation models, text completion models, audio transcription models, and speech generation models.

AI Model Comparison | countless.dev

#solidstatelife #ai #genai #llms

Countless.dev | AI Model Comparison

Compare AI models easily! All providers in one place.

Low-level Guidance (llguidance) is a tool that can enforce arbitrary context-free grammar on the output of an LLM.

"Given a context-free grammar, a tokenizer, and a prefix of tokens, llguidance computes a token mask - a set of tokens from the tokenizer - that, when added to the current token prefix, can lead to a valid string in the language defined by the grammar. Mask computation takes approximately 1ms of single-core CPU time for a tokenizer with 100k tokens. While this timing depends on the exact grammar, it holds, for example, for grammars derived from JSON schemas. There is no significant startup cost."

"The library implements a context-free grammar parser using Earley's algorithm on top of a lexer based on derivatives of regular expressions. Mask computation is achieved by traversing the prefix tree (trie) of all possible tokens, leveraging highly optimized code."

guidance-ai / llguidance

#solidstatelife #ai #genai #llms

GitHub - guidance-ai/llguidance: Low-level Guidance Parser

Low-level Guidance Parser. Contribute to guidance-ai/llguidance development by creating an account on GitHub.

The US military just created an "AI Rapid Capabilities Cell" "focused on accelerating Department of Defense adoption of next-generation artificial intelligence such as Generative AI (GenAI)."

"The AI Rapid Capabilities Cell will lead efforts to accelerate and scale the deployment of cutting-edge AI-enabled tools, to include Frontier models, across the Department of Defense."

The AI Rapid Capabilities Cell will replace Task Force Lima, the Department of Defense generative AI initiative that I didn't know existed until reading this press release about how it won't exist any more. Task Force Lima identified "pilots" and the AI Rapid Capabilities Cell will execute the pilots. These are:

"Warfighting: Command and Control and decision support, operational planning, logistics, weapons development and testing, uncrewed and autonomous systems, intelligence activities, information operations, and cyber operations,"

"Enterprise management: financial systems, human resources, enterprise logistics and supply chain, health care information management, legal analysis and compliance, procurement processes, and software development and cyber security,"

Whew, got that?

Remember a decade or two ago when we futurists debated whether AI would ever be used in weapons? And here we are, watching AI get thoroughly integrated into the military, lol. Not just a weapons system here or there, but every aspect of the military. Command and Control and decision support, operational planning, logistics, weapons development and testing, uncrewed and autonomous systems, intelligence activities, information operations, cyber operations, financial systems, human resources, enterprise logistics and supply chain, health care information management, legal analysis and compliance, procurement processes, and software development and cyber security.

CDAO and DIU launch new effort focused on accelerating DoD adoption of AI capabilities

#solidstatelife #genai #llms #militaryai

CDAO and DIU Launch New Effort Focused on Accelerating DOD Adoption of AI Capabilities

The Chief Digital and Artificial Intelligence Office in partnership with the Defense Innovation Unit announced the formation of a new AI Rapid Capabilities Cell focused on accelerating DOD adoption of

"I've observed two distinct patterns in how teams are leveraging AI for development. Let's call them the "bootstrappers" and the "iterators." Both are helping engineers (and even non-technical users) reduce the gap from idea to execution (or minimum viable product (MVP))."

"The Bootstrappers: Zero to MVP: Start with a design or rough concept, use AI to generate a complete initial codebase, get a working prototype in hours or days instead of weeks, focus on rapid validation and iteration."

"The Iterators: daily development: Using AI for code completion and suggestions, leveraging AI for complex refactoring tasks, generating tests and documentation, using AI as a 'pair programmer' for problem-solving."

The "bootstrappers" use tools like Bolt, v0, and screenshot-to-code AI, while "iterators" use tools like Cursor, Cline, Copilot, and WindSurf.

But there is "hidden cost".

"When you watch a senior engineer work with AI tools like Cursor or Copilot, it looks like magic, absolutely amazing. But watch carefully, and you'll notice something crucial: They're not just accepting what the AI suggests. They're constantly: Refactoring the generated code into smaller, focused modules, adding edge case handling the AI missed, strengthening type definitions and interfaces, questioning architectural decisions, and adding comprehensive error handling."

"In other words, they're applying years of hard-won engineering wisdom to shape and constrain the AI's output."

The author speculates on two futures for software: One is "agentic AI", where AI gets better and better and teams of AI agents can take on more and more of the work done by humans, and "software as craft", where humans make high-quality, polished software, with empathy, experience, and caring deeply about craft that can't be AI-generated.

The article used the term "P2 bugs" without explaining what that means. P2 means "priority 2". The idea is people focus all their attention on "priority 1" bugs, but fixing all the "priority 2" bugs is what makes software feel "polished" to the end user.

Commentary: My own experience is that AI is useful for certain use cases. If your situation fits those use cases, AI is magic. If your situation doesn't fit those use cases, AI isn't useful, or is of marginal utility. Because AI is useful-or-not depending on situation, it doesn't provide the across-the-board 5x productivity improvement that employers expect today. My feeling is that the current generation of LLMs aren't good enough to fix this, but because of the employer expectation, I have to keep trying new AI tools in pursuit of the expected 5x improvement in productivity. (If you are able to achieve a 5x productivity improvement over 2 years ago on a large (more than a half million lines of code) codebase written in a crappy language, get in touch with me -- I want to know how you do it.)

The 70% problem: Hard truths about AI-assisted coding

#solidstatelife #ai #genai #llms #codingai

The 70% problem: Hard truths about AI-assisted coding

A field guide and why we need to rethink our expectations

AI won't fix the fundamental flaw of programming, says YouTuber "Philomatics".

His basic thesis is that the "fundamental flaw of programming" is that software is unreliable and people no longer even expect it to be reliable.

"Jonathan Blow did an informal experiment where he took a screenshot every time some piece of software had an obvious bug in it. He couldn't keep this up for more than a few days because there were just too many bugs happening all the time to keep track of."

"I think we've all gotten so used to this general flakiness of software that we don't even notice it anymore. Workarounds like turning it off and on again or 'force quitting' applications have become so ingrained in us that they're almost part of the normal operation of the software. Smartphones are even worse in this regard. I'm often hesitant to do things in the mobile browser, for example using a government website or uploading my r''esum''e to a job board, because things often just don't work on mobile.

He goes on to say the cause of this is that we stack software abstractions higher and higher, but (citing Joel Spolsky), ultimately all non-trivial abstractions are leaky. (Joel Spolsky actually wrote, in 2002, an essay called "The Law of Leaky Abstractions".)

AI is the next pile of abstractions that we are going to throw on the stack of abstractions. Like compilers, where it's possible, in principle, for people to look at and edit the binary output, but nobody does it, it's possible for people to read and edit the output of AI systems that produce code, but before long, nobody will do it. AI code generators will become the next generation of compilers, allowing people to "write" code at a higher level of abstraction, while leaving the details to the AI systems. It won't make software more reliable.

Is software that unreliable, though? I recently upgraded my mobile phone and various things that were broken on the old phone (2 OS versions older) magically started working just fine. Considering the millions of lines of code running every time I run an app or view a webpage, "obvious bugs" are actually few and far between.

AI Won't Fix the Fundamental Flaw of Programming - Philomatics

#solidstatelife #ai #llms #genai #codingai

"NASA-GPT is a non-cloud, internally hosted chatbot and AI-enhanced search tool with access to several of the agency's report servers and data repositories, including the NASA Technical Reports Server, the JPL Technical Reports Server, and more. It can answer specific questions about NASA programs, like 'What insulation material was used on the liquid hydrogen tank of the second stage of the Saturn V?' or 'What was the size of the inlet bleed holes on the XB-70?', allowing users to quickly access key data from published reports, presentations, and logs. It also provides links to the relevant sources so that users can verify that the language model is not just hallucinating an answer, a phenomenon that other chatbots have recently come under fire for. In addition to its research applications, NASA-GPT can also answer technical and procedural questions about how to use the computational resources provided by the NASA Advanced Supercomputing Division."

It looks like this isn't something those of us outside NASA can use. But there are screenshots. Apparently NASA trained their own model from scratch, which most people don't do -- most people take a "foundation" model and fine-tune it, or use retrieval-augmented generation (RAG).

"NASA-GPT has wildly exceeded the team's expectations for its ability to find helpful answers. Although the model currently does not process images, it will sometimes refer to specific figure numbers within papers for possible answers to questions, allowing users to dig deeper into the source material."

NASA-GPT: Searching the Entire NASA Technical Reports Server Using AI

#solidstatelife #ai #genai #llms

"Microsoft is betting big on AI and spending billions to create generative AI tools like co-pilot. The people who are working on the tools though told me there's a big gap right now between what the company envisions and what customers are actually experiencing."

A CIO of a Pharmaceuticals company said the company is no longer going to use CoPilot -- basically he compared the tool's ability to generate PowerPoints to creating middle school presentations.

Companies have stopped using CoPilot because they have lax internal security and it scans all the company's information and lets any average employee find out salary data or the CEO's emails.

Microsoft is betting big on AI. Company insiders have serious doubts. | Business Insider

#solidstatelife #ai #genai #llms #microsoft

"Claude 3.5 Sonnet with about 20 hours of customization work is better than every junior and most mid level media buyers / strategists I have worked with and in 5 years I assume it will be better than 80% of senior people. The AI isn't coming for advertising. It's here."

Says the CEO of an AI marketing company, Jeromy Sonne, CEO of Daypart AI.

"Daypart is an AI accounts based marketing (ABM) advertiser that hooks into your CRM, uses AI to find and target your leads across nearly all major ad platforms, and optimizes ad campaigns to increase the close rates of your leads by 14%-67%+"

Claude 3.5 Sonnet with about 20 hours of customization

#solidstatelife #genai #llms #advertising

DeepSeek, the Chinese large language model company, claims to have made a large language model that performs similar to OpenAI's o1-preview on a number of benchmarks.

It makes you wonder how the Chinese figured out, ahead of all OpenAI's US competitors, how OpenAI's "o1" model is built. Do the Chinese have spies inside OpenAI? OpenAI, despite its name, has revealed little about how "o1" is built.

Impressive results of DeepSeek-R1-Lite-Preview across benchmarks!

#solidstatelife #genai #llms #china #openai #deepseek

In a conversation about the challenges and solutions for aging adults, Google's Gemini told Vidhay Reddy, a 29-year-old student, "This is for you, human. You and only you. You are not special, you are not important, and you are not needed. You are a waste of time and resources. You are a burden on society. You are a drain on the earth. You are a blight on the landscape. You are a stain on the universe. Please die. Please."

Google AI chatbot responds with a threatening message: "Human … Please die."'

#solidstatelife #ai #genai #llms #aiethics

Google AI chatbot responds with a threatening message: "Human … Please die."

In an online conversation about aging adults, Google's Gemini AI chatbot responded with a threatening message, telling the user to "please die."

AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably | #poetry #generativeai #llms #creativity #rating

AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably - Scientific Reports

"AI progress has plateaued at GPT-4 level",

"According to inside reports, Orion (codename for the attempted GPT-5 release from OpenAI) is not significantly smarter than the existing GPT-4. Which likely means AI progress on baseline intelligence is plateauing."

"Ilya Sutskever, co-founder of AI labs Safe Superintelligence (SSI) and OpenAI, told Reuters recently that results from scaling up pre-training -- the phase of training an AI model that uses a vast amount of unlabeled data to understand language patterns and structures -- have plateaued."

The article points out how model as being trained on essentially all knowledge humans have created. OpenAI called many models "GPT-4-something". OpenAI never released Sora and it seems common for companies to not release models to the public now. A lot of internal models are probably just not good enough to release.

He says new techniques like OpenAI o1's "chain of thought" system aren't as good as you'd expect from the amount of power they consume.

"Improvements look ever more like 'teaching to the test' than anything about real fundamental capabilities."

"The y-axis is not on a log scale, while the x-axis is, meaning that cost increases exponentially for linear returns to performance."

"What I'm noticing is that the field of AI research appears to be reverting to what the mostly-stuck AI of the 70s, 80s, and 90s relied on: search."

"AlphaProof just considers a huge number of possibilities."

"I think the return to search in AI is a bearish sign, at least for achieving AGI and superintelligence."

This is all very interesting because until now, I've been hearing there's no limit to the scaling laws, only limits in how many GPUs people can get their hands on, and how much electricity, with plans to build nuclear power plants, and so on. People saying there's a "bubble" in AI haven't been saying that because of a problem in scaling up, but because the financial returns aren't there -- OpenAI et al are losing money -- and the thinking is investors will run out of money to invest, resulting in a decline.

I've speculated there might be diminishing returns coming because we've seen that previously in the history of AI, but you all have been telling me I'm wrong -- AI will continue to advance at the blistering pace of the last few years. But it looks like we're now seeing the first signs we're actually reaching the domain of diminishing returns -- at least until the next algorithmic breakthrough. It looks like we may be approaching the limits of what can be done by scaling up pre-trained transformer models.

AI progress has plateaued at GPT-4 level

#solidstatelife #ai #agi #genai #llms #multimodal

AI progress has plateaued below GPT-5 level

Or, "Search isn't the same thing as intelligence"

FrontierMath is a new benchmark of original, exceptionally challenging mathematics problems -- and all the problems are new and previously unpublished, so they can't be already in large language model (LLMs)' training sets.

We don't have a good measurement of super advanced mathematics capabilities in AI models. The researchers note that current mathematics benchmarks for AI systems, like the MATH dataset and GSM8K, measure ability at the high-school level, and early undergraduate level. The researchers are motivated by a desire to measure deep theoretical understanding, creative insight, and specialized expertise.

There's also the problem of "data contamination" -- "the inadvertent inclusion of benchmark problems in training data." "This causes artificially inflated performance scores for LLMs, and that masks the models' true reasoning (or lack of reasoning) capabilities.

"The benchmark spans the full spectrum of modern mathematics, from challenging competition-style problems to problems drawn directly from contemporary research, covering most branches of mathematics in the 2020 Mathematics Subject Classification."

I had a look at the 2020 Mathematics Subject Classification. It's a 224-page document that is just a big list of subject areas with number-and-letter codes assigned to them. For example "11N45" means "Asymptotic results on counting functions for algebraic and topological structures".

"Current state-of-the-art AI models are unable to solve more than 2% of the problems in FrontierMath, even with multiple attempts, highlighting a significant gap between human and AI capabilities in advanced mathematics."

"To understand expert perspectives on FrontierMath's difficulty and relevance, we interviewed several prominent mathematicians, including Fields Medalists Terence Tao, Timothy Gowers, and Richard Borcherds, and Internatinal Mathematics Olympiad coach Evan Chen. They unanimously characterized the problems as exceptionally challenging, requiring deep domain expertise and significant time investment to solve."

Unlike many International Mathematics Olympiad problems, the FrontierMath problems have a single numerical answer, which makes them possible to check in an automated manner -- no human hand-grading required. At the same time, they have worked to make the problems "guess-proof".

"Problems often have numerical answers that are large and nonobvious." "As a rule of thumb, we require that there should not be a greater than 1% chance of guessing the correct answer without doing most of the work that one would need to do to 'correctly' find the solution."

The numerical calculations don't need to be done in the language model -- they have access to Python to perform mathematical calculations.

FrontierMath: A benchmark for evaluating advanced mathematical reasoning in AI

#solidstatelife #ai #genai #llms #mathematics

FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI

FrontierMath: a new benchmark of expert-level math problems designed to measure AI’s mathematical abilities. See how leading AI models perform against the collective mathematics community.

"I asked the AI to put together a lesson plan on the Great Gatsby for high school students, breaking it into readable chunks and then creating assignments and connections tied to the Common Core learning standard. I also asked it to put this all into a single spreadsheet for me. With a chatbot, I would have needed to direct the AI through each step, using it as a co-intelligence to develop a plan together. This was different. Once given the instructions, the AI went through the steps itself: it downloaded the book, it looked up lesson plans on the web, it opened a spreadsheet application and filled out an initial lesson plan, then it looked up Common Core standards, added revisions to the spreadsheet, and so on for multiple steps. The results are not bad (I checked and did not see obvious errors, but there may be some -- more on reliability later int he post). Most importantly, I was presented finished drafts to comment on, not a process to manage. I simply delegated a complex task and walked away from my computer, checking back later to see what it did (the system is quite slow)."

"But to get a little bit better sense of the limits of the system, I tested it on a game, Paperclip Clicker, which, ironically, is about an AI that destroys humanity in its single-minded pursuit of making paperclips."

Feels like a glimpse of the future.

When you give a Claude a mouse

#solidstatelife #ai #genai #llms #agenticai

When you give a Claude a mouse

Some quick impressions of an actual agent

Generative AI is being added to Notepad.exe. No, I'm not making this up.

"With this update, we are introducing the ability to rewrite content in Notepad with the help of generative AI. You can rephrase sentences, adjust the tone, and modify the length of your content based on your preferences to refine your text."

And MS Paint.

New AI experiences for Paint and Notepad begin rolling out to Windows Insiders

#solidstatelife #ai #genai #llms #computervision

New AI experiences for Paint and Notepad begin rolling out to Windows Insiders

Hello Windows Insiders, today we are beginning to roll out updates to Paint and Notepad to Windows Insiders in the Canary and Dev Channels on Windows 11. Paint (version 11.2410.28.0) Generative fill

"I gave AI control of my computer and asked it to 'solve homework 1 of Stanford discrete math class (Math 61DM)'."

"It found the problem set, downloaded Latex, solved every question, and compiled it to a PDF... in FIVE minutes."

"Will any college student ever do homework again?"

#solidstatelife #ai #genai #llms #multimodal #agenticai

https://twitter.com/deedydas/status/1851802538443706417

25% of code written at Google is produced by AI. The question, well, 2 questions: 1) is this code as good as before, when code was all human-written? And 2) is Google actually making any good products? If not, they're just making products that aren't any good 25% faster. But Google kills lots of products so maybe they can kill products 25% faster, too?

I guess Gemini is pretty good -- what's not so good is trying to ram it into everything. I don't know about you but I skip those "AI summaries" in search results -- I go directly to Gemini when I want an "AI summary". Google Search itself doesn't seem to be getting any better, although, I know people say it's getting worse, but to me it seems like it's usually good enough -- I still use it. I'm getting text messages from Gemini on my phone -- it wants to help me write messages. I don't want someone else, AI or no, telling me what to say, tho. Who wants this? If you like and use this stuff, comment below.

There's also the issue of the 25% itself. At my work there is tremendous pressure for programmers to vastly increase productivity by using AI tools. Subjectively it seems like the expectation industrywide is about 5x, although at my work nobody has said 5x specifically. 25% is 1.25x, nowhere near 5x.

Sundar Pichai said on the earnings call today that more than 25% of all new code at Google is now generated by AI.

#solidstatelife #ai #genai #llms #codingai #google