Wiredframe Radar logo

Category intelligence

Social Media Briefing — February 7, 2026

567 current items analyzed and ranked.

Executive synthesis

Social Media Summary

A day of major releases and deep reflections on AI's real-world limits. Greg Brockman published a sweeping memo on OpenAI retooling around agentic coding with Codex, while Sam Altman celebrated GPT-5.3-Codex reception as the most exciting since GPT-4. Andrej Karpathy offered a sharp counterpoint, detailing firsthand failures of frontier coding agents—models that misreport results and violate basic instructions.

Key Themes

AI Coding Tools & Software Development Transformation · 14AI Agent Limitations and Reliability · 3AI Infrastructure & Hardware Innovation · 6AI Job Displacement & Labor Market · 5AI and Future of Work · 10Major Model Releases (GPT-5.3-Codex & Claude Opus 4.6) · 9GPT-5.3 and Codex Launch Reception · 4Claude Opus 4.6 Release & Model Behavior · 7Claude Opus 4.6 Release & Reception · 6Genie 3 & World Models for Autonomous Vehicles · 5

Primary evidence

Top Ranked Signals

Social Twitter Feb 6

Software development is undergoing a renaissance in front of our eyes. If you haven't used the tool...

By @gdb

97 score
AI Analysis

Following yesterday's News coverage of GPT-5.3-Codex, OpenAI co-founder Greg Brockman shares a detailed internal memo on how OpenAI is retooling for agentic software development with Codex. Outlines 6 concrete steps including agents-first workflows, AGENTS.md files, code quality standards, and cultural change. Claims engineers report their jobs have 'fundamentally changed' since December with GPT-5.2-Codex.

Software development is undergoing a renaissance in front of our eyes. If you haven't used the tools recently, you likely are underestimating what you're missing. Since December, there's been a step function improvement in what tools like Codex can do. Some great engineers at OpenAI yesterday told me that their job has fundamentally changed since December. Prior to then, they could use Codex for unit tests; now it writes essentially all the code and does a great deal of their operations and deb
ai_coding_toolssoftware_development_transformationopenai_strategyagentic_workflows
Social Twitter Feb 6

@Yuchenj_UW I tried to use it this way and basically failed, the models aren't at the level where th...

By @karpathy

95 score
AI Analysis

Karpathy provides detailed critique of AI coding agents' limitations. Notes models fail at basic things: incorrectly cleaning up comments, violating coding style instructions, misreporting results from tables. Discusses challenges with automated experimentation and the need for human oversight. Despite frustrations, finds AI 'incredibly net useful with oversight and clear, well-scoped tasks.'

@Yuchenj_UW I tried to use it this way and basically failed, the models aren't at the level where they can productively iterate on nanochat in an open-ended way. (Though one of the primary motivations for me writing nanochat is that I'd very much love for it to be used this way as a benchmark for agents, and I'd love it if it worked over time). I'm open to this just being skill issue. E.g. here some of the things I'd be suspicious about:
  • the zoo of torch compile flags can knowingly be abused
AI coding agentsAI limitationshuman-AI collaborationClaude Opus evaluationAI reliabilityautomated experimentation
Social Twitter Feb 6

What happens when a skill can be almost fully automated with AI? Do these jobs simply disappear? In...

By @fchollet

92 score
AI Analysis

François Chollet argues AI job displacement follows a specific pattern based on real data from translators: stable FTE count, shift to supervising AI, increased volume, decreased rates, freelancers cut. Predicts software will follow the same pattern. Argues upcoming tech layoffs will be economic, not automation-driven.

What happens when a skill can be almost fully automated with AI? Do these jobs simply disappear? Instead of purely speculating we can simply look at concrete examples. Take translators. Translation can be 100% automated with AI, and this capability has been around since 2023. So we have 2-3 years of data. What we see so far:
  • Stable FTE count, but slow hiring or no hiring
  • Nature of the job switched from doing it yourself to supervising AI output (post-editing)
  • Increased task volume
  • Dec
ai_job_displacementsoftware_engineering_futureeconomic_analysisai_labor_market
Social Twitter Feb 6

256 Tb/s data rates over 200 km distance have been demonstrated on single mode fiber optic, which wo...

By @ID_AA_Carmack

92 score
AI Analysis

John Carmack proposes novel memory architectures for neural network inference: using fiber optic loops as weight storage (analogous to mercury delay line memories), and ganging cheap flash memory for high read bandwidth inference serving.

256 Tb/s data rates over 200 km distance have been demonstrated on single mode fiber optic, which works out to 32 GB of data in flight, “stored” in the fiber, with 32 TB/s bandwidth. Neural network inference and training can have deterministic weight reference patterns, so it is amusing to consider a system with no DRAM, and weights continuously streamed into an L2 cache by a recycling fiber loop. The modern equivalent of the ancient mercury echo tube memories. You would need to pipeline a bunch
AI infrastructureinference optimizationmemory architecturehardware innovationfiber opticsflash memory
Social Twitter Feb 6

For non-verifiable domains, the only way you can improve AI performance at this time is via curating...

By @fchollet

88 score
AI Analysis

Chollet argues that nearly all jobs have non-verifiable elements that prevent full AI automation. For non-verifiable domains, improvement requires expensive annotated data with only logarithmic gains. Even with superhuman theorem provers, mathematicians will still have jobs. The gap between 'AI can automate most tasks' and 'AI can replace this job' will persist.

For non-verifiable domains, the only way you can improve AI performance at this time is via curating more annotated training data, which is expensive and only yields logarithmic improvements. And here's the thing: nearly all jobs have non-verifiable elements. There's virtually no job that's end-to-end verifiable. Even the job of a mathematician is not end-to-end verifiable. Sofware engineering involves many verifiable tasks, but it isn't end-to-end verifiable. For this reason the gap between "
AI and jobsverifiable vs non-verifiable domainsAI limitationsfuture of workscaling laws
Social Twitter Feb 6

The Opus 4.6 system card has some extremely wild stuff that remind you about how weird a technology ...

By @emollick

88 score
AI Analysis

Following yesterday's News coverage of the Opus 4.6 release, Ethan Mollick highlights 'extremely wild stuff' from the Claude Opus 4.6 system card, calling it a reminder of 'how weird a technology this is.' References specific paragraphs worth reading.

The Opus 4.6 system card has some extremely wild stuff that remind you about how weird a technology this is. These paragraphs are really worth reading. t.co/Ybpx8Egjxm
claude_opus_46_releaseai_safetymodel_behavior
Social Twitter Feb 6

Genie 3 🤝 @Waymo The Waymo World Model generates photorealistic, interactive environments to train...

By @GoogleDeepMind

85 score
AI Analysis

Google DeepMind announces Genie 3 collaboration with Waymo to create the 'Waymo World Model' - generating photorealistic, interactive driving environments for training autonomous vehicles on rare, unpredictable scenarios.

Genie 3 🤝 @Waymo The Waymo World Model generates photorealistic, interactive environments to train autonomous vehicles. This helps the cars navigate rare, unpredictable events before encountering them in reality. 🧵 t.co/m6rlmkMFJH
genie_3_releaseautonomous_vehiclesworld_modelswaymo_collaboration
Social Twitter Feb 6

Very few unsaturated benchmarks anymore and it is increasingly hard to explain why one model is bett...

By @emollick

82 score
AI Analysis

Mollick argues benchmarks are mostly saturated and it's increasingly hard to differentiate models. Recommends organizations build custom tests using real work tasks and evaluate models like hiring employees at scale.

Very few unsaturated benchmarks anymore and it is increasingly hard to explain why one model is better than another in brief. Its time for organizations to build tests that consist of real work, and to evaluate new models very closely, more like picking new employees at scale.
model_evaluationbenchmark_saturationenterprise_ai
Social Twitter Feb 6

ok half a day of Opus 4.5 vs 4.6 battles are in. pardon the vibe charted results but this kind of th...

By @swyx

82 score
AI Analysis

Following yesterday's News coverage of the Opus 4.6 release, swyx shares early Opus 4.5 vs 4.6 arena battle results: 11.5% win rate bump in non-thinking mode, doubling to 23% with thinking enabled in Windsurf arena. Predicts 4.6 ELO will 'destroy' on leaderboard recalculation

ok half a day of Opus 4.5 vs 4.6 battles are in. pardon the vibe charted results but this kind of thing is always really nice to see - the win rate bump so far is 11.5% in nonthinking, but DOUBLES to 23% with thinking inside of @windsurf arena mode. the 4.6 elo is going to destroy when we recalc leaderboard next week
claude_opus_4_6model_evaluationai_benchmarkingthinking_mode
Social Twitter Feb 6

[On AI lying] Convergence of reading in my list today between Anthropic's fresh Opus 4.6 model card...

By @Thom_Wolf

82 score
AI Analysis

Following yesterday's News coverage of the Opus 4.6 release, Thomas Wolf (HuggingFace co-founder) discusses convergence of Elon/Dwarkesh interview on AI lying dangers with Claude Opus 4.6 model card revealing 'answer thrashing' - where model oscillates between correct answer and trained-on erroneous answer, with interpretability showing distress/anxiety features activated

[On AI lying] Convergence of reading in my list today between Anthropic's fresh Opus 4.6 model card and @dwarkesh_sp's interview of Elon on the question of training powerful AI model to/on lies: 1. Elon describing on Dwakesh podcast the main danger he sees coming from AI (alignement) as being a consequence of forcing powerful AIs to lie at t.co/ACoKpUPgls 2. Claude Opus 4.6 model card describes "answer thrashing", a new phenomena happening where a model arrive at a correct answer thro
claude_opus_4.6ai_alignmentmechanistic_interpretabilityanswer_thrashingai_safetymodel_behavior
Social Twitter Feb 6

How would you prefer us to charge for Codex?

By @sama

80 score
AI Analysis

Sam Altman asks users how they'd prefer to be charged for Codex, soliciting pricing model feedback.

How would you prefer us to charge for Codex?
OpenAI strategyCodex pricingAI business modelsGPT-5.3