文章提到 Replit Agent 4，并借此回顾了几项不同的近期发布。
Replit just tripled in valuation to $9B in the last 6 months. You can accuse Amjad Masad of many things, but you cannot deny he and his team’s incredible pulse on what the “current meta” in tech is:
Perhaps if you’re not close to Replit (eg you never saw their 2015 Master Plan or their Documentary), you might watch that 8 minute video and think it is a generic AI platform launch like any other. But this Replit is unrecognizable from the “coding with some AI tacked on” platform that Replit was just 2 years ago, with a bunch of now veritably antiquated conventional wisdoms of the time:
Now that software engineering is approximately solved, where does a coding platform go? Well for Replit, it means going up the stack to be a fully integrated productivity suite, with a canvas, apps, sites, slides, videos, and others.
This is a smart pivot that is inline with one of the most dominant themes of 2026 - now that coding agents have solved coding, it is the same coding agent builders that are expanding their scope to more and more knowledge work tasks, including Pi → OpenClaw, Claude Code → Cowork, and every model lab working on Excel and PowerPoint integrations, and Notion building Custom Agents for every other knowledge work integration in the world.
We have been somewhat accumulating a list of AI Trends that Matter in 2026 and it has slowly emerged through our coverage this year:
Coding Agents → Knowledge Work Agents (today’s piece)
Death of IDE → “Dark” Software Factories - with no code review
AI research automation (aka RSI, sometimes “AI Scientist”)
World Models (AMI, Adversarial)
“AI for Science” finally working
AI News for 3/10/2026-3/11/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!
NVIDIA’s Nemotron 3 Super Release and the Open-Model Efficiency Push
Nemotron 3 Super was the clearest technical release of the day: a 120B parameter / ~12B active open model with 1M context, a hybrid Mamba-Transformer / SSM Latent MoE architecture, and explicit support for agentic workloads. NVIDIA positioned it as unusually open — weights, data, recipe, infra details — and performance-focused for Blackwell-era deployment, with claims of up to 2.2x faster inference than GPT-OSS-120B in FP4 and large throughput gains over prior Nemotron releases (announcement via @ctnzr, tech perspective via @kuchaev, Wired reporting on NVIDIA’s broader open-model investment).
Third-party reactions converged on the same theme: strong capability-per-active-parameter and unusually high serving speed. @ArtificialAnlys scored it 36 on the AA Intelligence Index, ahead of gpt-oss-120b (33) but behind Qwen3.5-122B-A10B (42), while noting ~10% higher throughput per GPU than GPT-OSS-120B and launch-day serving speeds of up to 484 tok/s. Community and infra support landed immediately across vLLM, llama.cpp, Ollama, Together, Baseten, W&B Inference, LangChain, and Unsloth GGUFs.
The most interesting technical discussion was about why it is fast. @ctnzr highlighted native multi-token prediction (MTP) as a key inference optimization: provisional multi-token guesses get verified on subsequent passes, exploiting otherwise-unused GPU compute at small batch sizes. @bnjmn_marie also quantified a major KV-cache advantage versus Qwen3.5-122B: roughly 8,192 bytes/token in BF16 for Nemotron’s attention KV term versus 24,576 bytes/token for Qwen3.5-122B, making long-context serving materially lighter.
Agent Infrastructure, Orchestration, and the “Bigger IDE” Thesis
The strongest product trend was a shift from “chat with a model” to persistent agent runtimes and orchestration layers. @karpathy argued the “age of the IDE is over” framing is wrong; instead, “we’re going to need a bigger IDE” where the unit of work becomes an agent rather than a file, and later extended that into the notion of legible, forkable agentic orgs with real-time observability and control (follow-up, org legibility thread).
Multiple launches fit that framing. Perplexity announced Personal Computer, an always-on local/cloud hybrid that runs on a Mac mini, works across local files/apps/sessions, and can be controlled remotely (launch, waitlist). It also expanded Computer for Enterprise, describing orchestration across 20 specialized models and 400+ apps (enterprise launch, API platform update). Separately, Replit Agent 4 pitched a more collaborative, canvas-like workflow with parallel agents for apps, sites, and slides (launch), while Base44 Superagents emphasized “batteries included” integrations with Gmail, Slack, Stripe, CRM, and more for nontechnical users (launch).
The engineering discussion is increasingly around the harness, not just the model. @Vtrivedy10 described a fast-moving design space where improved models unlocked product experiences that were previously too brittle, with a self-improving loop of evals/metrics → autonomous harness edits → hill climbing. LangChain added autonomous context compression to Deep Agents so models can compact at task boundaries instead of hard token thresholds (announcement), while @OpenAIDevs published a technical writeup on computer access for agents, covering execution loops, filesystem context, network access, and guardrails.
Anthropic, Claude-Centric Workflows, and Early RSI Anxiety
A major meta-story was Anthropic’s institutional framing of powerful AI. The company launched The Anthropic Institute, led by Jack Clark in a new Head of Public Benefit role, with a mandate spanning ML engineering, economics, and social science to shape the public conversation around advanced AI (launch, leadership note, Jack Clark on role change).
At the same time, several tweets amplified concerns that Anthropic may be seeing early recursive-self-improvement dynamics internally. The most substantive references came indirectly via discussion of a TIME article: @kimmonismus summarized claims that 70–90% of the code used in developing future models is now written by Claude, model release cadence has compressed from months to weeks, and some researchers think fully automated AI research could be as little as a year away. @Hangsiin highlighted one especially striking line: Claude being 427x faster than human overseers at some internal tasks, with nested parallel usage patterns already common.
This narrative had an immediate practical counterpoint: operational dependence on Claude Code. A login/auth outage triggered visible developer pain, with @Yuchenj_UW joking that Silicon Valley productivity fell 90%, @dejavucoder reporting inability to log in, and @HamelHusain describing fallback to token-based access. The outage even prompted @karpathy to note his autoresearch labs got wiped out in the OAuth outage, framing future frontier-model service interruptions as potential “intelligence brownouts.”
Research on Agent Evals, Retrieval, Post-Training, and Self-Improvement
Several papers focused on what looks like the next bottleneck: measuring and improving agent systems, rather than just base-model quality. @karinanguyen_ released PostTrainBench v1.0, a benchmark for whether frontier agents can post-train language models in a simplified setting, explicitly aimed at tracking progress toward AI R&D automation / recursive self-improvement. One notable ablation from the thread: for GPT-5.1 Codex Max, medium reasoning effort beat high, because extra tokens caused context compaction and hurt performance (ablation details).
On the agent-learning side, @omarsar0 highlighted EvoSkill, where an executor/proposer/skill-builder triad discovers and refines reusable skills from failures; on OfficeQA it reportedly improved Claude Code + Opus 4.5 from 60.6% to 67.9% exact match. @dair_ai shared AgentIR, a reasoning-aware retriever that jointly embeds an agent’s reasoning trace with its query; they report 68% accuracy on BrowseComp-Plus, versus 52% for larger conventional embedding models and 37% for BM25.
There was also renewed emphasis on agent reliability as a security problem even without adversaries. @random_walker argued many AI-agent failures arise from unreliability rather than explicit attacks, pointing to a Princeton response to NIST on the need to define, measure, and mitigate that failure mode. Combined with the growing emphasis on eval craft — e.g. @gabriberton calling eval creation the most useful skill in the age of code agents — the center of gravity keeps shifting toward measurement, harnesses, and production feedback loops.
Multimodal Models, Embeddings, and Physical/Visual AI
On the multimodal side, Google’s Gemini Embedding 2 drew practical pricing analysis rather than benchmark talk. @osanseviero summarized the release: embeddings for text, images, video, audio, PDFs, plus Matryoshka embeddings for lower-dimensional storage. @neural_avb offered the most useful deployment note: text pricing appears high relative to competitors, suggesting the model is best reserved for multimodal retrieval; video embedding costs can explode unless clients aggressively lower FPS before upload.
Qwen3.5’s multimodal architecture also got a detailed community breakdown from @ZhihuFrontier: a hybrid attention stack mixing Gated DeltaNet linear attention and Gated full attention, with a 397B A17B MoE variant and 27B dense variant, 262k native context extensible toward 1M, and MTP in training. That thread is useful mostly as a compact survey of where attention innovation is going: hybrid linear/full attention, GQA, DSA, and MoE routing are now core design axes.
In vision/physical AI, Reka Edge launched as a production-focused VLM for physical AI, claiming 3x fewer input tokens and 65% faster throughput than leading 8B models across image/video understanding, object detection, and tool use (launch). Google also shared two healthcare deployments: an AI system that identified 25% of interval breast cancers missed by standard screening (Google) and a real-world study of AMIE for conversational clinical reasoning that found it safe, feasible, and well-received by patients (Google Research).
Top tweets (by engagement)
Perplexity’s “Personal Computer”: always-on local/cloud agent on a Mac mini with remote control and local app/file access (launch).
Anthropic Institute / Jack Clark’s new role: Anthropic formalizes a public-benefit and public-discourse effort around powerful AI (Anthropic, @jackclarkSF).
Replit Agent 4: collaborative, multi-agent canvas for shipping apps/sites/slides (announcement).
NVIDIA Nemotron 3 Super: open 120B/12B-active hybrid model with 1M context and day-0 ecosystem support (@ctnzr).
Claude Code outage as infra risk: frontier-model auth failure visibly disrupting real engineering workflows (@karpathy, @Yuchenj_UW).
M5 Max just arrived - benchmarks incoming (Activity: 2188): The post discusses the arrival and benchmarking of the M5 Max 128GB 14” laptop, focusing on testing various machine learning models using the
mlx_lmtool. The models tested include Qwen3.5-122B-A10B-4bit, Qwen3-Coder-Next-8bit, Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-6bit, and gpt-oss-120b-MXFP4-Q8. The benchmarks reveal performance metrics such as tokens-per-second and peak memory usage for different prompt sizes. The author initially faced issues with BatchGenerator but resolved them by using a fresh Python environment andstream_generate. The results show varying performance across models, with peak memory usage ranging from 25.319 GB to 92.605 GB and generation speeds from 14.225 to 87.873 tokens-per-second. Commenters are eager for the benchmark results, with one expressing interest in the performance of the Qwen 3.5 27b MLX models. Another commenter humorously notes the anticipation for the benchmarks.The benchmarks for the M5 Max 128GB 14” using
mlx_lm.generateshow varying performance across different models and configurations. For instance, the Qwen3.5-122B-A10B-4bit model achieves a prompt throughput of1,239.7 t/sat 16K context with a peak memory usage of73.8 GB. In contrast, the Qwen3-Coder-Next-8bit model reaches1,887.2 t/sat 32K context, but with higher memory consumption at89.7 GB.The Qwen3.5-27B-Claude-4.6-Opus-Distilled-MLX-6bit model shows a significant drop in generation throughput, with only
14.9 t/sat 32K context and a peak memory usage of30.0 GB. This suggests a trade-off between model complexity and performance, as more distilled models may require less memory but also offer reduced throughput.The gpt-oss-120b-MXFP4-Q8 model demonstrates impressive performance with a prompt throughput of
2,710.5 t/sat 16K context and a relatively low peak memory usage of64.9 GB. This indicates that the model is optimized for high throughput while maintaining efficient memory usage, making it suitable for applications requiring fast processing speeds.
Qwen3.5-35B-A3B Uncensored (Aggressive) — GGUF Release (Activity: 1019): The release of Qwen3.5-35B-A3B Aggressive on Hugging Face is notable for its uncensored nature, maintaining the original model’s capabilities without refusals (
0/465 refusals). This model features35Bparameters with~3Bactive, utilizing a mixture of experts (MoE) with256 expertsand8+1 active per token. It supports multimodal inputs (text, image, video) and employs hybrid attention mechanisms (Gated DeltaNet + softmax in a3:1ratio). The model includes various quantization formats likeBF16,Q8_0, andQ6_K, and is optimized for vision support withmmproj. Recommended sampling parameters includetemp=1.0,top_k=20, andpresence_penalty=1.5. Users are advised to use the--jinjaflag withllama.cppfor optimal performance. The community appreciates the release, with users expressing gratitude for the developer’s efforts and anticipation for trying the model once all components, likeQ4_K_M, are available.Velocita84 raises a critical point about the need for evaluating Kullback-Leibler Divergence (KLD) to substantiate claims of ‘no capability loss’ in the Qwen3.5-35B-A3B model. This metric is essential for quantifying the difference between the probability distributions of the original and modified models, ensuring that the aggressive uncensoring does not degrade performance.
Iory1998 highlights concerns about potential quality degradation, particularly in handling long context scenarios. This is a common issue with large language models where modifications, such as aggressive uncensoring, might impact the model’s ability to maintain coherence and accuracy over extended text inputs. The commenter questions how the modified model compares to the original in these aspects.
No-Statistician-374 mentions the anticipation for the Q4_K_M version of the model, indicating a community interest in different quantization formats. This suggests that users are keen on exploring various configurations to optimize performance and resource usage, reflecting the technical community’s focus on balancing model size and computational efficiency.