Hello everyone-<\/p>\n
Could is right here! Spring time sunshine (hopefully) and a gradual stream of financial institution holidays (within the UK a minimum of\u2026) so loads of time to atone for your Knowledge Science and AI studying.<\/span>Numerous nice content material beneath and I actually encourage you to learn on, however listed below are the edited highlights if you’re quick for time! <\/span><\/p>\n Following is the newest version of our Knowledge Science and AI publication. Hopefully some attention-grabbing subjects and titbits to feed your information science curiosity. NOTE: If the e-mail doesn\u2019t show correctly (it’s fairly lengthy\u2026) click on on the \u201cOpen On-line\u201d hyperlink on the prime proper. Suggestions welcome!<\/p>\n useful new fast hyperlinks: Numerous thrilling information science and AI happening, as all the time!<\/p>\n Bias, ethics, range and regulation proceed to be sizzling subjects in information science and AI…<\/p>\n Whether or not and tips on how to regulate AI stays a sizzling matter<\/p>\n Certainly, the AI cellphone has triggered an uproar in China. Inside days, a lot of China\u2019s greatest apps blocked the Doubao cellphone. They noticed it as a critical threat to information safety. Constructed into the working system of the cellphone itself, it has a sort of grasp key that offers the embedded AI agent blanket entry to the display screen, all app content material, and the flexibility to faucet or click on as if it had been a consumer. Critics dubbed the agent a \u201cburglar\u201d with \u201cgod\u2019s fingertips\u201d rising dangers of malicious enter and intrusion assaults by legal actors. For the banks, it was not possible to tell apart actions taken by the agent and people of the consumer, creating myriad vulnerabilities for fraud and hacking.However on Monday, Gov. Gavin Newsom of California, a Democrat, defied Mr. Trump by issuing an government order that requires security and privateness guardrails for A.I. corporations contracting with the state. He additionally mentioned he would battle to protect California\u2019s legal guidelines that present safeguards towards A.I.-related catastrophic harms, scams and dangers for kids.<\/p>\n Prefer it or not, AI retains impacting increasingly more areas of enterprise and society<\/p>\n AI is clearly getting used to aim to sway elections- <\/span>Lots of of Pretend Professional-Trump Avatars Emerge on Social Media<\/p>\n Well being is likely one of the hottest subjects researched on chatbots, however how correct are they of their responses?<\/span> It will depend on what context they’re given it appears.<\/span><\/p>\n How a lot affect is AI going to have on economies?<\/span> Nobody is bound, however some suppose the affect shall be very giant<\/p>\n Good comparability right here with what occurred to journey brokers with the appearance of the internet- <\/span>A case research in job displacement<\/p>\n So the decline of journey brokers was pushed largely by technological displacement. AI is a really totally different sort of shock from the web of the Nineteen Nineties, and its results on labor markets might grow to be fairly totally different too. However the story of the journey agent continues to be instructive, as a result of it demonstrates what occurs to staff when a expertise renders their occupation out of date. Because it seems, the story shouldn’t be solely bleak.<\/p>\n And with regards to utilizing AI, <\/span>extra analysis is displaying that the quick time period achieve may nicely come on the expense of long run hurt<\/span><\/p>\n In the meantime, specialists have gotten more and more frightened about cybersecurity dangers<\/p>\n Mythos Preview has already discovered 1000’s of high-severity vulnerabilities, together with some in each main working system and internet browser. Given the speed of AI progress, it won’t be lengthy earlier than such capabilities proliferate, doubtlessly past actors who’re dedicated to deploying them safely. The fallout\u2014for economies, public security, and nationwide safety\u2014might be extreme. Undertaking Glasswing is an pressing try and put these capabilities to work for defensive functions.<\/p>\n Lastly\u2026 the robots are getting higher! <\/span>Robots beat human data at Beijing half-marathon<\/p>\n As all the time, a lot of new developments on the analysis entrance and loads of arXiv papers to learn…<\/p>\n Good to see some attention-grabbing analysis from outdoors of the large industrial labs<\/p>\n First there’s <\/span>Aurora from Collectively.ai: bettering speculative decoding- the place a small, quick \u201cdraft\u201d mannequin is used to foretell a number of future tokens, which a bigger \u201cgoal\u201d mannequin verifies in parallel <\/span><\/p>\n Speculative decoding goes stale in manufacturing \u2014 draft fashions can drift and offline retraining cannot all the time hold tempo. After all Deep Thoughts is fairly irrepressible on the analysis entrance:<\/p>\n A brand new strategy to image-text encoding: <\/span>TIPSv2<\/p>\n And a extra philiosophical piece: <\/span>The Abstraction Fallacy: Why AI Can Simulate However Not Instantiate Consciousness (thanks Dirk!)- wants a quiet room and a while, however well worth the effort!<\/span><\/p>\n The framework proposed right here explicitly separates simulation (behavioral mimicry pushed by car causality) from instantiation (intrinsic bodily structure pushed by content material causality). Establishing this ontological boundary reveals why algorithmic image manipulation is structurally incapable of instantiating expertise.<\/p>\n Anthropic continues do wonderful work in interpretability, in search of buildings within the weights that appear to embody particular ideas: <\/span>Emotion ideas and their operate in a big language mannequin<\/p>\n How steady are fashionable LLMs- wonderful work from Nvidia in figuring out how <\/span>modifications in sure key weights can change outputs dramatically<\/p>\n Deep neural networks are susceptible to catastrophic failure from flipping only a few signal bits in mannequin parameters. We current Deep Neural Lesion (DNL), a data-free methodology that identifies and exploits crucial parameters throughout imaginative and prescient and language domains. Thought scary and a bit disconcerting: once we take into consideration \u201cChain-of-thought\u201d we assume the mannequin first causes a plan, after which follows the plan to the reply. What if it was the opposite approach round? <\/span>Due to this fact I’m. I Suppose (I did test and it\u2019s not an April Idiot!)<\/span><\/p>\n On this paper, we current proof that detectable, early-encoded selections form chain-of-thought in reasoning fashions. Particularly, we present {that a} easy linear probe efficiently decodes tool-calling selections from pre-generation activations with very excessive confidence, and in some instances, even earlier than a single reasoning token is produced.<\/p>\n Making use of the LLM ideas to totally different information units<\/p>\n Very cool: CLIP (encoding photos in the identical vector area as textual content), however for geospatial information: <\/span>UniGeoCLIP – outcomes look spectacular on geospatial issues<\/span><\/p>\n Should Learn- a superb instance of making use of the ideas of LLMs to several types of information: <\/span>PRAGMA: Revolut Basis Mannequin<\/p>\n Our strategy pre-trains a Transformer-based structure with masked modelling on a large-scale, heterogeneous banking occasion corpus utilizing a self-supervised goal tailor-made to the discrete, variable-length nature of economic data. The ensuing mannequin helps a variety of downstream duties corresponding to credit score scoring, fraud detection, and lifelong worth prediction: robust efficiency will be achieved by coaching a easy linear mannequin on prime of the extracted embeddings and will be additional improved with light-weight fine-tuning. Via in depth analysis on downstream duties, we display that PRAGMA achieves superior efficiency throughout a number of domains immediately from uncooked occasion sequences, offering a general-purpose illustration layer for monetary functions.<\/p>\n After which there\u2019s drumming! <\/span>DexDrummer: In-Hand, Contact-Wealthy, and Lengthy-Horizon Dexterous Robotic Drumming<\/p>\n Developments within the wild world of huge language fashions, basis fashions, world fashions…<\/p>\n Earlier than we delve into the corporate updates, I believed it could be helpful to take a look at a few key themes throughout the large three: Safety, Coding, and Agent infrastructure<\/p>\n Again on the round-up entrance, OpenAI had a busy month:<\/p>\n Google is so spectacular at progressing capabilities throughout an enormous vary:<\/p>\n Anthropic is unquestionably making waves<\/p>\n Meta is doubtlessly again within the recreation, with the primary launch from their expensively assembled Superintelligence Labs – <\/span>Muse Spark<\/p>\n And even Microsoft is releasing: <\/span>3 new world class MAI fashions, obtainable in Foundry<\/p>\n On the Open Supply entrance:<\/p>\n The Chinese language contingent proceed to impress- <\/p>\n And the novelty of a brand new open supply mannequin from a non-Chinese language lab: <\/span>Trinity-Massive-Considering from Arcee within the US (why not Europe?)<\/span><\/p>\n Numerous functions and suggestions and methods this month<\/p>\n First on the appliance aspect:<\/p>\n If you’re within the enterprise of writing analysis papers, <\/span>PaperOrchestra may be price taking a look at<\/span><\/p>\n Writing a analysis paper is brutal. Even after the experiments are accomplished, a researcher nonetheless faces weeks of translating messy lab notes, scattered outcomes tables, and half-formed concepts into a elegant, logically coherent manuscript formatted exactly to a convention\u2019s specs. For a lot of recent researchers, that translation work is the place papers go to die.<\/p>\n A crew at Google Cloud AI Analysis suggest \u2018PaperOrchestra\u2018, a multi-agent system that autonomously converts unstructured pre-writing supplies \u2014 a tough concept abstract and uncooked experimental logs \u2014 right into a submission-ready LaTeX manuscript, full with a literature overview, generated figures, and API-verified citations.Walrus was educated on 19 totally different bodily eventualities spanning 63 bodily variables in each 2 and 3D. Walrus makes use of new instruments for adaptive computation and improved stability as a way to obtain correct long-term rollouts whereas co-adapting sampling and distribution to enhance coaching throughput regardless of dealing with various dimensions, resolutions, and facet ratios.<\/p>\n Study when you code: <\/span>your private coding tutor in Google Colab<\/p>\n Placing AI to good use- monitoring authoritarian regimes: <\/span>PropagandaScope<\/p>\n I looked for it in PropagandaScope. The outcome confirmed what I had suspected however couldn’t show. State media had been saturating minority areas with this phrase for years. Xinjiang Day by day used it at 17.4 instances the speed of Folks\u2019s Day by day. Tibet Day by day, 12.1 instances. Ningxia Day by day, 10.3 instances.<\/p>\n The locations with the very best ethnic minority populations acquired the heaviest propaganda. The locations China calls \u201cautonomous\u201d had the least autonomy over their very own newspapers.<\/p>\n Helpful – <\/span>Google quietly launched an AI dictation app that works offline<\/p>\n Woah – <\/span>Coaching mRNA Language Fashions Throughout 25 Species for $165<\/p>\n We constructed an end-to-end protein AI pipeline overlaying construction prediction, sequence design, and codon optimization. After evaluating a number of transformer architectures for codon-level language modeling, CodonRoBERTa-large-v2 emerged because the clear winner with a perplexity of 4.10 and a Spearman CAI correlation of 0.40, considerably outperforming ModernBERT. We then scaled to 25 species, educated 4 manufacturing fashions in 55 GPU-hours, and constructed a species-conditioned system that no different open-source venture presents. <\/p>\n Numerous nice tutorials and howto guides:<\/p>\n Coding Brokers are develop into very highly effective, and nicely price exploring. Just a few pointers:<\/p>\n So, in brief, we are able to summarize it like this: What’s an agent harness I hear you ask? <\/span>The Anatomy of an Agent Harness <\/p>\n I nonetheless discover Reinforcement Studying a bit mystical if I\u2019m honest- I discovered this to be a <\/span>helpful abstract for the way you may use it for LLM Brokers<\/p>\n I\u2019ve come throughout<\/span> DSPy quite a lot of instances, and actually wish to use it and like it- however nonetheless havnt fairly discovered the proper use case. I discovered this to be precise instance of its implementation: <\/span>How we optimized Sprint\u2019s relevance choose with DSPy<\/p>\n DSPy lets you outline that setup\u2014job, information, and metric\u2014after which systematically seek for immediate variants that enhance efficiency on that metric. We used DSPy\u2019s GEPA optimizer (a way that iteratively improves prompts by analyzing the place the mannequin disagrees with people and producing suggestions) to adapt and optimize the relevance-judging program for a selected goal mannequin\u2014on this case, gpt-oss-120b.I wished to know how transformers truly route data between layers. Not on the degree of \u201cconsideration computes weighted averages,\u201d however on the degree of what bodily occurs to a tensor because it strikes by the community. What will get preserved, what will get overwritten, and why.<\/p>\n A video helped construct a few of the preliminary instinct. From there I picked up the paper, Consideration Residuals (Kimi Group, 2026), and determined to reimplement it from scratch. No HuggingFace, no pre-built transformer blocks. Simply torch.einsum, nn.Parameter, and a toy dataset sufficiently small that I may hint each matrix multiplication by hand.accuracy \/\u02c8\u00e6k.j\u0259.r\u0259.si\/ noun 1. The diploma to which predictions match actuality inside a subset of the info chosen to yield the very best doable accuracy. 2. Whichever of the true constructive fee and the true unfavourable fee is probably the most spectacular.<\/p>\n drive analytics, ML and AI into manufacturing<\/p>\n Raft is so fetch: The Raft Consensus Algorithm defined by \u201cImply Ladies\u201d<\/p>\n Raft is a consensus algorithm utilized in distributed methods to make sure that information is replicated safely and persistently. That sentence alone will be complicated. Hopefully the analogy on this publish may also help folks perceive the way it works. In honor of nationwide Imply Ladies day (\u201con October third he requested me what day it was\u201d), I current the Raft Consensus Algorithm as defined by the film Imply Ladies. <\/p>\n Quantization from the bottom up<\/p>\n Qwen-3-Coder-Subsequent is an 80 billion parameter mannequin 159.4GB in dimension. That is roughly how a lot RAM you would wish to run it, and that is earlier than excited about lengthy context home windows. This isn’t thought of a giant mannequin. Rumors have it that frontier fashions have over 1 trillion parameters, which might require a minimum of 2TB of RAM. The final time I noticed that a lot RAM in a single machine was by no means.<\/p>\n However what if I instructed you we are able to make LLMs 4x smaller and 2x quicker, sufficient to run very succesful fashions in your laptop computer, all whereas dropping solely 5-10% accuracy.<\/p>\n That is the magic of quantization.<\/p>\n Holding it easy – <\/span>The Git Instructions I Run Earlier than Studying Any Code<\/p>\n Columnar Storage is Normalization<\/p>\n A unique approach to consider columnarization like that is that it is akin to very excessive kind of database normalization.<\/p>\n As an alternative of 1 vast desk that is represented by a bunch of vectors of information, you may consider columnar information as a set of tables which all have a main key plus one further attribute<\/p>\n Helpful tutorial: <\/span> Construct a Scalable Serverless Social Media Ingestion & Analytics Pipeline on AWS<\/p>\n If it’s worthwhile to host a prototype and even one thing a lot bigger, Vercel is defintiely price trying out. Their choices for AI functions are significantly helpful: <\/span>Agentic Infrastructure<\/p>\n As the ultimate actor shifts from human to machine, infrastructure has to adapt once more. It has to work for software program that acts on behalf of customers, writes itself, and more and more wants to know its personal conduct in manufacturing. This new technology of agentic software program calls for Agentic Infrastructure.<\/p>\n It\u2019s not one evolution, however three: Longer thought scary reads – sit again and pour a drink! … <\/p>\n Andrej Karpathy on the <\/span>hole in understanding of AI functionality<\/p>\n The primary concern I believe is round recency and tier of use. I believe lots of people tried the free tier of ChatGPT someplace final yr and allowed it to tell their views on AI slightly an excessive amount of. It is a group of reactions laughing at numerous quirks of the fashions, hallucinations, and many others. Sure I additionally noticed the viral movies of OpenAI’s Superior Voice mode fumbling easy queries like “ought to I drive or stroll to the carwash”. The factor is that these free and previous\/deprecated fashions do not mirror the aptitude within the newest spherical of state-of-the-art agentic fashions of this yr, particularly OpenAI Codex and Claude Code.<\/p>\n The ying and yang of Anthropic\u2019s Claude Mythos!<\/p>\n The Mythos Threshold – Joe Reis<\/span><\/p>\n It was probably the most accountable, rigorously staged rollout of a harmful functionality that the business had ever seen.<\/p>\n And it was terrifying. A general-purpose mannequin \u2014 one no person educated for cybersecurity \u2014 was autonomously discovering, chaining collectively, and exploiting vulnerabilities at a degree exceeding most human safety researchers. The apparent query hung over every part: what else can it try this we haven\u2019t thought to check but?<\/p>\n The discourse lasted a few week. The market dipped. It recovered. That is the way it works now.<\/p>\n Claude Mythos #2: Cybersecurity and Undertaking Glasswing – Zvi Mowshowitz<\/span><\/p>\n As you take into account all of this, don’t forget that Mythos is a big step in direction of automated AI R&D and sufficiently superior AI, and likewise reveals some shadows of what such a future AI shall be able to doing. We’re headed into existential hazard, along with the very actual catastrophic cybersecurity threats we have to sort out now.<\/p>\n Anthropic\u2019s Claude Mythos Launch Is Constructed on Misinformation – Devansh<\/span><\/p>\n The issue is every part round it. Virtually each main outlet or commentator overlaying Mythos labored from Anthropic\u2019s press supplies and never the precise main sources such because the CVE advisories, the exploit code, the 44-prompt transcript, the 244-page system card. Whenever you learn them and add the AISLE replication research, the pink crew writeups, the Glasswing accomplice agreements, Anthropic\u2019s personal decpetive framings, and a really totally different image emerges: certainly one of misinformation and hype.<\/p>\n The Lacking Layer: Why Your AI Agent Fails \u2014 and What Truly Fixes It – Ben Lorica<\/span><\/p>\n Software program engineering groups had been the primary to hit this wall at scale, and their response presents a sensible blueprint for each area now constructing AI-powered functions. Their conclusion was counterintuitive: scaling AI reliably shouldn’t be primarily about making the underlying mannequin smarter. It requires a totally totally different self-discipline centered on constructing a structured, automated atmosphere across the mannequin. That self-discipline known as harness engineering, and its ideas lengthen nicely past writing software program.<\/p>\n The AI Labs Have A $7 Doritos Drawback – Vin Vashishta<\/span><\/p>\n Doritos costs jumped practically 50% between 2021 and early 2026. Some baggage crossed $7, which is quite a bit for junk meals. Walmart instructed PepsiCo to chop costs for over a yr. PepsiCo tried every part besides reducing costs, from promotions and shrinkflation to new product traces. None of it labored. Income turned unfavourable for the primary time in over a decade. Walmart pulled shelf area and handed it to Takis and its personal personal label.<\/p>\n There are structural parallels to AI. OpenAI, Anthropic, Google, and Microsoft are all navigating the identical query Frito-Lay didn’t reply. Shoppers and enterprises are evaluating AI subscriptions the identical approach consumers evaluated a $7 bag of chips. Is that this price it? Is that this one thing I would like, or one thing I can skip?<\/p>\n The AlphaFold second for supplies isn’t any time quickly – Connor Blake<\/span><\/p>\n One central query was \u201cwhat’s the AlphaFold of supplies science, and the way can we get it in 6-24 months?\u201d That is an ill-posed query as a result of proteins have a bunch of good properties that supplies don\u2019t have. Particularly, amino acid sequences are principally enough to find out construction in physiological situations of curiosity, however composition and even lattice unit cells usually are not enough to find out construction of Actual Bodily Supplies. Regardless of what everybody needs to imagine primarily based on customary bulk modeling methods, most supplies are disordered and have complicated interfaces, and should you can\u2019t seize the implications of those, fashions will output rubbish.<\/p>\n The Revenge of the Knowledge Scientist – Hamel Husain<\/span><\/p>\n For years, delivery AI meant maintaining information scientists and MLEs on the crucial path. With LLMs, this stopped being the default. Basis-model APIs now permit groups to combine AI independently.<\/p>\n Getting lower out of the loop rattled information scientists and MLEs I do know. If the corporate not wants you to ship AI, it’s truthful to wonder if the job nonetheless has the identical upside. The harsher story folks inform themselves: except you’re pretraining at a foundation-model lab, you aren’t the place the motion is.<\/p>\n I learn it the opposite approach. Coaching fashions was by no means many of the job. The majority of the work is establishing experiments to check how nicely the AI generalizes to unseen information, debugging stochastic methods, and designing good metrics. Calling an LLM over an API doesn’t make this work go away.<\/p>\n We\u2019re truly working out of benchmarks to higher sure AI capabilities – Lawrence Chan<\/span><\/p>\n
\n<\/span>committee; <\/span>ethics; <\/span>analysis; <\/span>generative ai; <\/span>functions and howtoguides; <\/span>engineering_tips; <\/span>massive image concepts; <\/span>enjoyable; <\/span>reader updates; <\/span>jobs<\/p><\/blockquote>\n![\"\"](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!STbg!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F23200e7a-5491-45ff-bc22-cfa7e5424ded_1554x873.png\")
<\/div>\n<\/figure>\n<\/div>\n![\"A](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!MXBy!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F128f282e-f37b-402d-a28a-08372658916c_1600x1316.png\" "\\"A")
<\/div>\n<\/figure>\n<\/div>\n![\"\"](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!GLWC!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F17b63128-39da-43be-86f6-326e53c6b1b9_1197x1347.png\")
<\/div>\n<\/figure>\n<\/div>\n![\"Solve](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!MIkn!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F11a641e4-8e3b-455b-821b-45fc0fbfe9fc_1560x1080.png\" "\\"Solve")
<\/div>\n<\/figure>\n<\/div>\n![\"The](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!3iFG!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1ee77ad2-6ea2-4bab-ba55-8da82dd10c5c_1024x667.jpeg\" "\\"The")
<\/div>\n<\/figure>\n<\/div>\n
\nAurora fixes this. It is an open-source, RL-based framework that learns immediately from stay inference traces and repeatedly updates the speculator with out interrupting serving.
\nKey outcomes:
\n\u2192 Actual-time adaptation throughout shifting site visitors domains
\n\u2192 1.25x further speedup over a well-trained static speculator
\nThe headline discovering: on-line coaching from scratch can outperform a rigorously pretrained static baseline.<\/p>\n![\"TIPSv2](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!i3Ki!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff962854a-6283-4fbc-b2f0-b618b5836cd4_1555x1306.jpeg\" "\\"TIPSv2")
<\/div>\n<\/figure>\n<\/div>\n![\"Emotion](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!nEkY!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb44ddf5f-a02e-4761-ba97-33e32c6b1a02_3764x2380.webp\" "\\"Emotion")
<\/div>\n<\/figure>\n<\/div>\n
\nOur strategy requires solely write entry to saved weights\u2014no coaching information, no optimization, minimal computation. This makes it sensible beneath life like risk fashions the place attackers compromise mannequin storage by firmware exploits, rootkits, DMA assaults, or Rowhammer vulnerabilities.
\n– ResNet-50: 2 signal flips \u2192 99.8% accuracy drop
\n– Masks R-CNN \/ YOLOv8-seg: 1\u20132 flips collapse detection and segmentation
\n– Qwen3-30B & Nemotron 8B: Few flips cut back reasoning and job accuracy to near-zero<\/p>\n![\"UniGeoCLIP](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!vnFF!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3e7cfa39-0ffe-4683-96d1-f4a5c6aa013d_1243x505.png\" "\\"UniGeoCLIP")
<\/div>\n<\/figure>\n<\/div>\n![\"Dimensional](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!BNoK!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F054d7dc7-d28b-44d5-9164-44d859f8e4ea_2974x852.png\" "\\"Dimensional")
<\/div>\n<\/figure>\n<\/div>\n![\"\"](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!bZnc!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc46c23fa-1738-43d2-b784-e6d20894e601_653x335.png\")
<\/div>\n<\/figure>\n<\/div>\n![\"\"](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!CCZN!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa7a5c903-df9d-4b72-82a3-075d09339ed8_2400x1350.jpeg\")
<\/div>\n<\/figure>\n<\/div>\n![\"\"](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!dMMH!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdc00dd2f-d946-41ef-a123-40817e9812ee_1432x1080.png\")
<\/div>\n<\/figure>\n<\/div>\n
\n– LLM: the uncooked mannequin
\n– Reasoning mannequin: an LLM optimized to output intermediate reasoning traces and to confirm itself extra
\n– Agent: a loop that makes use of a mannequin plus instruments, reminiscence, and atmosphere suggestions
\n– Agent harness: the software program scaffold round an agent that manages context, software use, prompts, state, and management movement
\n– Coding harness: a particular case of an agent harness; i.e., a task-specific harness for software program engineering that manages code context, instruments, execution, and iterative suggestions<\/p>\n![\"Image\"](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!7lNx!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4a07b679-cb29-42d2-9349-33d1951c38ef_680x407.jpeg\" "\\"Image\\"")
<\/div>\n<\/figure>\n<\/div>\n![\"The](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!bt-v!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc0626449-a101-4e82-bbe7-4698c7b56096_820x469.png\" "\\"The")
<\/div>\n<\/figure>\n<\/div>\n![\"The](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!f2ZT!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F007492b5-1143-4057-a8af-490f80ce27ed_1500x760.webp\" "\\"The")
<\/div>\n<\/figure>\n<\/div>\n![\"\"](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!0_OR!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3cef16c3-eda1-4ca3-90bc-9957bb446ffb_826x593.png\")
<\/div>\n<\/figure>\n<\/div>\n
\n– Infrastructure for coding brokers to deploy to
\n– Infrastructure for constructing and working brokers
\n– Infrastructure that itself is agentic<\/p>\n
![\"\"](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!qVk3!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F66945ef2-d610-4d06-b7c1-2d9bb64f03d7_678x486.png\")
<\/div>\n<\/figure>\n<\/div>\n![\"Image\"](\"https:\/\/substackcdn.com\/image\/fetch\/$s_!yClN!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd052e0ae-5993-4dab-b51a-59504f85d4ac_1448x1086.jpeg\" "\\"Image\\"")
<\/div>\n<\/figure>\n<\/div>\n