Tesla FSD 14.3.2 Has a Problem

Key Points

▶ 0:46 Chris used Grok built into the car to plan a six-point route through downtown Ann Arbor for FSD testing, only adding the first destination himself.
▶ 1:23 He immediately flagged a flaw in Grok's route—several destinations were clustered together—and Grok responded by canceling the old route and setting a new one.
▶ 1:47 The wake word feature still isn't available and there's no "spring update" yet, but Chris predicts that once it arrives you'll just tell the car where to go and confirm—and that confirmation button may eventually disappear.
▶ 2:12 FSD behaviors are converging rapidly, observed while testing in downtown conditions with red-light running still exposed.
▶ 2:26 FSD 14.3.2 has finally rolled out wide, with "pretty much everyone" receiving the update.
▶ 2:37 The only notable exclusion is users on the spring update track, who do not have it yet.
▶ 2:41 FSD 14.3.2 still has many lingering issues, though some behaviors (like slowing for a small speed bump) remain good decisions.
▶ 2:52 Regressions are surprising because 14.2 was considered excellent for ~97% of driving, with only minor known shortcomings.
▶ 3:45 Overall, 14.3 is not bad but seems slightly worse than 14.2, despite expectations that it would feel "super sentient."
▶ 3:51 FSD 14.3.2’s brake stabbing is slightly worse than 14.2 but far better than 14.1.
▶ 4:04 A bird flying in front of the car triggers a sudden, harsh brake stop, which the narrator finds funny and excessive.
▶ 4:57 The braking wasn’t actually needed—the bird was going to make it across safely—and this overreaction is exactly why they’re testing live.
▶ 5:22 Major software rewrites like FSD 14.3.2 naturally come with regressions, and adjustments are expected.
▶ 5:40 The disengagement menu misbehaved by shifting the entire visualization over the map and blocking the interface until the user selected an option.
▶ 5:57 Tesla quickly fixed the issue by moving the menu to the lower corner, letting drivers continue using other controls without being blocked.
▶ 6:06 The on-screen FSD prompt has changed from taking over the whole screen to being placed "out of the way."
▶ 6:11 The prompt still remains until the driver makes a selection, requiring the user to choose a preference to proceed.
▶ 6:20 The driver gives positive feedback, noting the UI updated very quickly and calling the change "nice to see."
▶ 6:22 A pedestrian is nearly hit by a bus, which the narrator calls "crazy."
▶ 6:25 The pedestrian is jaywalking only ~5 feet from a crosswalk, which the narrator finds ironic and funny.
▶ 6:37 New FSD version has "brake stabbing" issues and many complaints, even though the car handled a pedestrian well without unnecessary braking.
▶ 6:53 Wide rollouts are valuable because thousands of users provide feedback and clips, offering a fuller picture than one person testing alone.
▶ 7:17 Negative feedback is strong enough that many users are asking for FSD 14.2 back, which the driver isn't surprised by since 14.2 was smoother.
▶ 7:41 Faster reaction times are apparent in this build, but this speed is linked to the stabbiness, jerkiness, and non-smoothness he’s seeing—arguing it needs to be smoothed out.
▶ 7:55 He highlights the release note about Tesla rewriting the AI compiler and runtime from the ground up, yielding a 20% faster reaction time—calling it “amazing” and noting improved model iteration speed.
▶ 8:17 He connects that iteration speed to the rapid succession of early 14.x updates, and at ▶ 8:36 recalls early 14.2’s brake-stabbing being “unacceptable” at stop signs, right turns, and around pedestrians—until it was eventually smoothed out completely.
▶ 8:57 FSD attempts to stop in a no-parking zone, and the driver immediately flags it as illegal parking.
▶ 9:08 The driver decides to report the incident through the FSD feedback system, then continue the trip.
▶ 9:26 The driver clarifies he did not leave the car parked there and moved on; the onlooker was likely staring because of earlier camera setup, not the FSD behavior.
▶ 9:26 Briefly notes they did not park in that spot and moved on, transitioning to discussion of the FSD update.
▶ 9:30 The update intentionally reduces faster reaction times to create a smoother driving experience, and this trade-off worked.
▶ 9:38 Reaction times are not as good as the previous version but remain "really good" compared to version 13.
▶ 9:52 Brake stabbing has returned in the current software, a central regression for users.
▶ 9:54 Tesla claims faster model iteration, but users aren't seeing the benefits.
▶ 10:05 Despite rapid internal version updates, improvements are not reaching customers—a key disconnect in the update cadence.
▶ 10:08 Tesla benefits from faster model iteration speed and has rolled out the software to many users.
▶ 10:14 Literal driverless robo-taxis are operating in three separate cities with dozens of vehicles, functioning essentially all the time.
▶ 10:41 Tesla is moving at a decent speed compared to Waymo, and the speaker approves of the robo-taxi rollout.
▶ 10:56 FSD 14.3.2 has been rolled out to basically everyone, yet weird brake stabbing issues still persist in this wide release.
▶ 11:04 The system illegally parked despite a very clear no-parking sign, showing failure to obey unambiguous traffic controls.
▶ 11:11 The car misread a diamond-shaped "prepare to stop ahead" sign as a "do not enter" sign, prompting an unnecessary U-turn — revealing flawed sign interpretation.
▶ 11:41 Earlier v14 versions were impressively smooth, e.g., FSD correctly predicted another car would move out of the way.
▶ 11:49 Current driving performance remains "super impressive," with roughly 95–99% of behavior being "wickedly impressive."
▶ 12:02 The system is "super safe," with no recallable safety-related incidents across any v14 version.
▶ 12:14 Most FSD disengagements are due to speed control and awkward full stops at stop signs, not safety issues.
▶ 12:33 The speaker emphasizes they have not had a single safety-related takeover, despite taking over at times.
▶ 12:36 A minor "brake jam" is highlighted using the visualization's blue path, showing an abrupt braking flaw rather than a serious incident.
▶ 12:41 FSD's harsh braking "doesn't feel good" to ride in, which is a critical trust problem now that "random normal people" are using it—if they don't trust it, they won't use it.
▶ 12:58 Despite the comfort issues, FSD is "way safer than your average person," with a "minuscule" chance of an accident compared to a typical driver.
▶ 13:16 Any FSD accidents are likely to be "really low speed," and the speaker is waiting to see whether one happens at scale, adding "hopefully not."
▶ 13:27 With millions of cars on the road, accidents are statistically inevitable.
▶ 13:33 The speaker spots police activity ahead, shifting from a general safety point to a live real-world situation.
▶ 13:33 The thought about the police activity is cut off mid-sentence, leaving the observation incomplete.
▶ 13:41 Speed control in Hurry Mode is described as "really bad" due to inconsistency.
▶ 13:52 On a 55 mph road, the car sometimes drove 72–75 mph, which the speaker accepts as expected for Hurry Mode.
▶ 14:04 On the same road type, other times it drove only 50–52 mph, creating unpredictable behavior and frustration.
▶ 14:10 Even the most aggressive speed mode (Mad Max) fails to fix the core issue: a fundamental lack of consistency across FSD speed modes.
▶ 14:17 Example of inconsistency: on a 65 mph road, Standard mode drives at 72 mph (ticket-worthy) while Chill mode drives at 67 mph (too slow for traffic).
▶ 14:27 The main user frustration is the inability to simply fit in with traffic, leaving drivers either awkwardly speeding or lagging behind with no reliable middle-ground option.
▶ 14:36 The speaker is surprised by the recent lack of software updates, calling it a notable issue at the time of recording.
▶ 14:38 Despite the update drought, the software is still described as "incredibly impressive" and remains the primary tool for driving.
▶ 14:45 Vehicle data shows 96% of all driving was done using NFSD (self-driving), with the speaker estimating their personal usage at 98% and the rest from Stephanie's trips.
▶ 15:04 The car remains "super impressive" overall despite odd maneuvers, and the speaker plans to continue looking for positive moments.
▶ 15:12 Online feedback is deeply contradictory—some users complain heavily, while others claim zero interventions; the speaker is skeptical of both extremes.
▶ 15:37 A post claims Zach completed a coast-to-coast cannonball run on this FSD version, but the speaker doubts it is possible given the version's frequent brake interventions and quirks.
▶ 15:42 The speaker is skeptical of a reported cross-country FSD run, citing likely issues like brake stabbing, bad map data causing odd navigation, and speed compliance problems.
▶ 16:07 They speculate the driver may have hidden manual intervention—like pressing the accelerator or using a slower mode—since no video was posted, which raises the question of whether it still counts as a success.
▶ 16:21 The speaker avoids taking a strong position on the debate and decides to move on with the drive.
▶ 16:30 Spot a shiny black-wrapped Cybertruck (Cyberbeast) with a University of Michigan sticker at Ann Arbor City Hall, calling it "kind of crazy."
▶ 16:43 Expresses surprise at seeing it there and compliments the owner: "If that's yours, nice job. I like it. That thing looks sick."
▶ 16:54 Ends with a wistful aside: "I miss my Cyber Truck. I should..."
▶ 17:36 FSD exits the tight parking lot "absolutely perfect": it never fully stops or brakes hard, instead slowly creeping out due to obstructed visibility from parked cars.
▶ 17:47 The cautious, continuous movement is praised as human-like—slow enough to stop if someone appears, but not frozen or indecisive—earning explicit approval from the narrator.
▶ 18:03 FSD approaches a left turn on a one-way street where it must yield to all oncoming traffic, setting up a right-of-way judgment that the transcript cuts off mid-description.
▶ 18:19 Driver asks in-car Grok to find top 2-3 complaints about FSD 14.3.2 from recent X posts mentioning the exact version.
▶ 18:51 Grok returns top complaints: harsh micro braking, messed-up lane changes, navigation glitches, plus hesitations and construction-zone confusion.
▶ 19:13 Driver reacts positively, saying he needs to use this real-time search in videos and notes Grok excels at X content—and was even able to pull Reddit results.
▶ 19:27 FSD goes through a yellow light instead of braking, which the driver notes is unusual for the system.
▶ 19:35 The vehicle makes an unexpected maneuver, and the driver initially questions it before accepting and letting FSD proceed.
▶ 19:50 The driver credits Grok for setting the destination, showing AI’s role in route decisions.
▶ 19:55 The driver drops his earlier skepticism as FSD legitimately handles a pedestrian encounter.
▶ 19:59 FSD brakes as a pedestrian walks past, then proceeds, earning the driver's approving laugh.
▶ 20:07 FSD aborts a tight fit maneuver, and at ▶ 20:10 the driver praises it: "Great job, FSD. This is great."
▶ 20:24 The narrator admits that, as a human driver, they would not have chosen the route the AI took.
▶ 20:31 They don't strongly criticize the AI, noting the system fully rescued itself from the situation.
▶ 20:37 The vehicle parks in a 15-minute limit spot, which the narrator accepts for a short stop.
▶ 20:48 The drive reaches its final destination, with the narrator declaring "win" and praising the Grok route setup for taking them to never-before-visited locations untested with FSD.
▶ 21:00 The narrator reveals they intentionally told Grok to "take me through places that will be difficult for the car," deliberately stress-testing FSD rather than taking a casual route.
▶ 21:06 The first notable mistake on the Grok-planned route is flagged: "there you go. You kind of saw a mistake there," though its exact nature is not yet explained.
▶ 21:15 A second pedestrian event begins with a harsh driver reaction ("Whoa, whoa, whoa"), immediately judged as "so bad."
▶ 21:25 The initial avoidance maneuver is praised as "perfect" and a "great job," but FSD then follows up with an unnecessary brake tap and a harsh braking/swerving overcorrection at ▶ 21:30-▶ 21:38, even after the pedestrian was already leaving.
▶ 21:50 The driver has mixed feelings—acknowledging the system's intent to be safe ("It's moving away from a person") but ultimately expressing frustration ("I do hate it") as the section ends in exasperated laughter.
▶ 22:13 Grok arrives at its final destination; the driver finds no parking options selected and decides to let FSD handle it.
▶ 22:27 The drive is summarized as zero takeovers, zero disengagements, and zero interventions, with mistakes happening too quickly to practically intervene.
▶ 23:09 The driver compares it to a robotaxi scenario—nobody would take over—and downplays rear-ending concerns because events occur fast, especially in a slow downtown area.

Summary

Executive Summary

Key Points

Video Sections

Exact Transcript